Akio Arakawa - Session II

Edwards:

Okay. We’re back with Professor Arakawa on July 18th, and he’s just showing me an IBM research report which is the first part of the thick volume on numerical simulation of weather and climate. This one is called Numerical Experiments with the Mintz-Arakawa General Circulation Model — Part 1: Description of the Model, also by W. E. Langlois and H. C. W. Kwok. The report number is RJ-501, dated May 24th, 1968.

Arakawa:

And this is a description of not the very first 2-level model.

Edwards:

That’s right, that’s what you said yesterday.

Arakawa:

I told you that but I was wrong. This is the description of the early stage of this model.

Edwards:

All right, so the second 2-level model.

Arakawa:

Yes. So the Rand report and this report describe the same model, but this represents the early stage of this.

Edwards:

Yeah, okay. Thank you.

Arakawa:

I found a copy of the proposal.

Edwards:

Aha, great. This is a National Science Foundation proposal in 1970.

Arakawa:

Right. It’s a very short proposal, but that was apparently enough to get a level of sophistication [chuckles]. That doesn’t say much about what the proposal was, but it does describe what we have done, and so I think this may be helpful for the record.

Edwards:

Yes, great, thanks so much. Anything else from yesterday that you wanted to say?

Arakawa:

Not right now. I don’t recall.

Edwards:

All right. The first thing I wanted to ask you before we leave models and go on to some other things is tell me as much as you can remember about as you moved from one model to the next in this process, to what degree did you recode the model entirely from scratch, and to what degree did you incorporate previous models into each successive generation?

Arakawa:

We have never recoded the model, and but now we are developing the completely new model from scratch. But for these models, just we kept additions and divisions and so on, so even though they not the same, we find very, very old statements.

Edwards:

Interesting. Do you have any idea how much of the original, the earliest model still remains in the present form?

Arakawa:

You mean right now? [Yes.] Well, very few. Oh of course the main details that changed are of the physics and the mathematics of the model, but also the programmers are constantly working on it to increase the efficiency and readability of the code. So although I said there may be some old statements, but well, I guess it’s very different right now. But the name of the code, some valuables can still remain, the same name of the subroutine, remain the same and so on. But we are right now developing a new GCN from scratch, so in parallel we have something. We have a version of this — that is still our standard GCN, but from the scratch we are building a brand new model.

Edwards:

Okay. Tell me about the newest one: what’s unique about it and what are its features?

Arakawa:

We try to make it everything unique — new. The mathematical structure is completely different. It is based on what we call generalized particle coordinate. Standard coordinate used the GCN and the NW models are called a sigma coordinate. It’s basically pressure. Pressure decreases in height, so we can use that as a coordinate. Basically pressure, but no matter by itself is pressure, okay, so the surface sigma called 1. So the surface is coordinated surface. That gives the term this convenience. Then there is mountain, then the surface is coordinated, then above with that they coordinate that. That is sigma coordinates — standard coordinate form. But those coordinator surfaces are not necessarily close to the surface of motion. Okay, suppose that coordinate surface, if there is a mountain like that, the coordinator surface above still try to follow mountain, but the motion may be just one period for weather. And also since that surface is somewhat influenced by the steep topography, some computational difficulties appear and so on. And there’s almost consensus that it’s good to use entropy at that part called particle coordinate. Entropy is constant if there is no heating. That means if you use entropy at coordinate surface, then the air stays at the same surface. But that surface can move up and down, but as the air moves up and down. So as far as there is no heating, there is no flow crossing that surface. That makes many things very clean. They had to be free from computational error. I think everybody agrees to that advantage, but everybody recognized the difficulty with variable it took air to come down, and the coordinator surface intersects at the surface like that. Here we are working on that kind of model with that kind of coordinates, but eventually we generalize it near the surface, just like a sigma coordinate which follows air to the surface, but almost immediately away from the surface it becomes the entropy coordinate surface.

Edwards:

Okay, so the surface coordinate system is still sigma, but in the next level up it becomes entropy coordinate system.

Arakawa:

Right, with a smooth transition between two coordinates. But that raises many difficulties. Just a paper came out just recently on that kind of model. And that problem was of that model is a very good, but so far basically we had the dynamics. We have the physics, boundary physics, and we have standard condensation process like that, okay. But essentially this is one new model. As we put the addition of processes, we had to seriously think about this, and what’s a good way of doing based on the various many years’ experience, and so we tried to put in something new and good, but not necessarily too complicated. So it takes time; we just slowly moving.

Edwards:

Is this model going to be for a parallel processing machine, or are you still using single processors?

Arakawa:

Not this new model. These models are done with parallel processing machine, but of course the new model should be recorded.

Edwards:

Okay. Now, another question I had was, so does this mean that all of the models that you have done use a finite difference scheme?

Arakawa:

Yes.

Edwards:

So you haven’t gone to the spectral methods. Why?

Arakawa:

Because I don’t see much advantage of that. I shouldn’t say I object to the spectral method. A main advantage of the spectral method is related to the program near the pole, that recorder is very small. Unless we use a special technique, the time interval has to be very short. And the pole is at a similar point of spherical coordinates. Now all of those kinds of difficulties do not exist in spherical coordinates, or it’s easy to overcome. The standard technique to avoid using too short data piece is called a semi-implicit scheme, and actually make the certain terms responsible for the fastest moving wave, treated implicitly. Suppose we try to predict from time N to N+1, then formulate those terms based on the varies at N+1, which is unknown. So we need the iteration what the matter is converging, but then if we do that then we can use longer ΔT. Okay? Now, it’s a bit complicated procedure if we stay with the finite difference, but it’s very straightforward when we use that aspect or method. Okay. So then major advantage, at least one of the major advantages of this particular method comes from that. So using longer ΔT works so the model becomes more efficient. However, I heard some conceptual argument that the spectral method tried to look at just a certain field as a superposition wave. Now suppose we are interested in the gradient right over Los Angeles. Then the spectral method would look at that as a superposition of many waves. Okay? There may not be gradient-zero — but the spectral method would look at the zero as a superposition of many waves. Well, if the result is zero, just the waves cancel out. Then you calculate the gradient for each wave, and that’s very straightforward, its wave, then the superimpose it.

Edwards:

I see. Okay.

Arakawa:

So it’s a tremendously complicated way of producing zeros. And that requires — suppose we are interested in the gradient say right over Los Angeles, we need information over the entire globe. [Laughs] So the concept is kind of ridiculous, but of course there is a problem associated with such a ridiculous method, and therefore you can put the error in the one thing and the one part can be in principal influenced immediately, instantaneous, and they work.

Edwards:

Interesting.

Arakawa:

And the finite difference method is a very broad method, and it can include the spectral method at one extreme. The more accurate, or using more and more grid points we can in principle increase accuracy. Both we use at the grid points and globe, then the highest accuracy possible, then they apply the spectral method. And they are together very flexible. That means very easy to make mistakes. And what the spectral method cannot — so, they’re very flexible, so while the spectral method is very straightforward, basically the only decision you have to make is where to truncate. Of course you can wash out the waves. But one of the differences is there are so many possibilities. So some people hated that flexibility, but I like that flexibility, so that’s more potential. But that means we have to be careful because it can be very bad. The spectral method that is usually more economical than the usual standard of finite difference scheme, well usually. But when we increase the degree of freedom more or the number of grid points, that is definitely the direction we are going, then the cross point advantage of spectral method in terms of the economy will eventually be lost. With the first method amount of computation is near in proportion now, good, and number of grid points where the spectral method the distinctions are different.

Edwards:

Geometric, yeah.

Arakawa:

I don’t remember exactly, but, yeah.

Edwards:

Okay. Okay, let me look at my questions here. Here’s one thing we didn’t talk about yesterday. There are two basic kinds of instability in models: physical instability related to incorrect parameterizations and so on, and computational instabilities. And this is one of the things that you’ve worked on extensively, especially in your early work. I’d just like to hear you talk about that a little bit, especially in terms of what the relationship is proportionally between problems with physical instabilities and problems with computational instabilities in models.

Arakawa:

Oh boy. There are many kinds of computational instabilities, and of course you know many kinds of physical instabilities. Just last April I gave about a four hour lecture in Korea on instability mechanisms in the atmosphere and the atmospheric models, and so it’s a survey of the adjustment mechanisms in the atmosphere, the different quantities, variables made to adjust each other, or the atmosphere itself will adjust to the surface conditions, like that. And there are many kinds of adjustment mechanisms, but sometimes they fail to adjust due to the existence of instabilities. I reviewed the several kinds of instabilities due to the failure of the adjustment and its consequence and so on, and then went on how those physical instabilities can be disposed of in certain models. And that is best expanded version in my paper “Adjustment Mechanisms in Atmospheric Models” that was just recently published.

Edwards:

Okay. Well I can look at that.

Arakawa:

That is the Japan Meteorological Society. That I have, only different. Now this was the lecture originally. Originally, based on the lecture I gave at the International Physics Symposium on Data Assimilation in Meteorology and Oceanography held March 1995. Here, this is a new developing field, and there is some concern about the inflow to this field from outside of meteorology in [???] and so on. So, I was asked to give a lecture on the atmospheric models. Well, without assuming much pre-knowledge, but in there I emphasized adjustment mechanisms because those who want to analyze the data should know what kind of adjustment mechanisms exist, or may not exist.

Edwards:

This is the paper called “Adjustment Mechanisms in Atmospheric Models.”

Arakawa:

That’s right, yes. I integrated that to my lectures and the paper presented at that Symposium. The reason I brought up it was my lecture at the Korean Meteorological Society at the University where I expanded adjustment and instability mechanisms in the atmosphere and atmospheric models. Okay, so that includes the computational instability? Not much that we know, but of course it reviews the physical instabilities. So now your question is how I…?

Edwards:

What’s the relationship between them? And I’m especially interested in how important computational instability is relative to physical instabilities and how that’s changed over time.

Arakawa:

Computational stabilities necessary condition. Of course and there are other, more than one way of defining stability, but usually the most simpleminded definition of stability is model runs without blowing up. Okay? And that’s a necessary condition. But usually not far from sufficient conditions — model may keep running, but producing garbage. That is stable, and a definition of stability which is a pretty common definition. So what I have been advocating is computational instability is of course too bad, but that is not enough to think. What we did with discrete system, we discretized. Finite difference, spectral, that really doesn’t matter. But that original system is continuous system, because in that way we can tremendously distort the behavioral solution. And just instability is just an example, one bad example. But even a stable behavioral solution can be very different, and disparate case, even there is a mode which doesn’t have a counterpart in the continuous system of complete studious [?] mode. And so one thing I am advocating is we have to maintain the analogy of the discrete system to the continuous system, and just the [???] program just a small portion of it. My 1966 paper, the motivation of that paper was to overcome practical problem of nonlinear computation instability, a relatively newly found computational instability at that time, and so it is nonlinear so it’s difficult to get the mathematical analysis of the nature of the instability. Linear instability is relatively easy. The nonlinear instability was a new kind of an instability discovered originally by the Phillips experiment in 1956, because his model blew up. The usual computational instability is eliminated or reduced by making this short ΔT. That of course he did, but it didn’t help; it can only just slightly delay the catastrophe. So that there is a new kind of computational instability Phillips a few years later called the nonlinear computational instability. In my 1960 paper, the motivation is to overcome nonlinear instability so that we can construct a GCM. But the other point, that paper, it’s more than that. It’s really not a pure mathematician’s approach. Really I think I’m a kind of physicist, physicist approach, just this physical behavioral of discrete system should be analogous to that of the original system. So that is the basic, but in addition to that there are really good physical instabilities. Of course in nature, the instability can exist, but if the [???] grows due to physical instability, then it’s now relative they were equilibrated. That’s why nature doesn’t blow up. Or instability may exist, but may not be actually realized, depending on the scale or the other conditions, and instability may exist, potential for instability may exist but it may not be realized. But for a model to smooth a severe instability is something very easy to happen in a model. So my overall idea is not really the distinguished [???], just try to have some discrete system — we have to have a discrete system so that we can compute — discrete system which is physically analogous to the continuous system.

Edwards:

Right. To what degree do you feel that the… I guess I’d want to say computational instabilities in the models have been reduced to insignificance at this point. Are there still major areas, major issues, for example truncation errors and things of that sort?

Arakawa:

The idea with the truncation error, it’s a bit old idea, and I shouldn’t say unimportant, but that doesn’t tell the whole story at all. And what I think is very important is the computational mode. It’s a mode which does not have a counterpart to the original equation. So it’s spurious, but we cannot talk about the error. The error is the difference of that solution from the true solution, but there is no true solution corresponding to that. So that for that kind of mode, the concept of truncation error is useless. So now that computational mode usually appears with very small spatial structure, either horizontally or vertically, or it can be in time actually. That thing was first found in time, but now it is in space, it can be a space in the vertical. And that appears with the small scale. But when the model doesn’t have much physics, especially when model doesn’t have moisture, it may not matter so much and the model result may have a noise we can smooth just cosmetically. But if there is a moisture, then condensation, precipitation may appear after responding to that small scale. And whenever condensation is involved, it is subject to rectification. Suppose… [Draws a wave form on a piece of paper] Usually condensation appears when the vertical velocity is apart. Then the air too. Suppose vertical velocities are like that. This is a time or it can be space.

Edwards:

Up and down, according to the wave. Changing vertical velocity, positive and negative.

Arakawa:

Right. Then the condensation appears when the vertical velocity is upward, but nothing happens here.

Edwards:

I see. All right, so we get a series of sort of lumps of condensation with no condensation on the downward side of the wave.

Arakawa:

This is another plus to the rectification of the active kind from AC to DC. Now, each of these may have the shorter time scale or the shorter space scale. We may be interested in a much larger scale, but if this kind of situation happens, this just adds up. So that’s another way that the cumulus parameterization is important. This may be the of cumulus, but that adds up. Now, suppose this kind of situation takes place and is associated with computational mode, which is a kind of zigzag structure. Then that’s horrible. It adds up. So it’s a nonlinear interaction. But usually nonlinearity we are aware, with others in that case and in fluid dynamics in general, is due to advection and velocity times the gradient of that variable. That’s nonlinear. Nonlinearity we are familiar with. That’s in the form of product — quadratic. But this is not that kind of nonlinearity; this is a more like not quadratic, more on and off type, recognized sign. And the nonlinearity must have the handle. In my mind, I don’t distinguish too much of the computational on the physical. I’ll just say I’m concerned with the physics of the discrete system — it can be computationally distorted. But pure computational instability is just a very small portion with it.

Edwards:

Well, this sort of leads naturally into another question, which is a whole area we haven’t talked about at all yet, which is validation of your models. When we were talking yesterday in the article you wrote about Jule Mintz, you talked about how important observation was to him, and your early background was also in observational meteorology. So clearly both of you were strongly aware of the real atmosphere, and you have just been talking about how difficult it is to model in a discrete form, so I’m curious about to what degree and when you began to compare the model results and how with what sorts of data. Part of what I’m interested in here is that during this whole period of your modeling efforts, the amounts of data that became available were increasingly large, a very rapid, upward curve, and especially in the seventies, when satellite data became available. So I just want to hear about how that influenced your work, if it did.

Arakawa:

I have to be a bit careful [laughs]. Of course we’ve got to know how nature behaves to model nature intelligently, of course, so modeling has not become a pure mathematical task. Really, and I think I quoted at the opening in that short lecture, and also Rossby had the same feeling, that we don’t know the basic equations. Okay. Many people, meteorologists, don’t recognize that. [Static on recording begins] Okay, we know that the load increases, and loading involved a molecular viscosity and a molecular heat conduction, and those processes are only for the scale [???].

Edwards:

Right. Molecular scale processes.

Arakawa:

Right. And it is a molecular scale… [Retrieves chart]. This is the horizontal scale. And in this one we are interested in this scale.

Edwards:

The planetary and synoptic scale.

Arakawa:

…in this spectral range, 100 to 10,000 kilometers. And we know pretty well the…

Edwards:

The one millimeter to one centimeter, which is on this chart.

Arakawa:

But that. Oh, of course that is nice, but that is not all.

Edwards:

Right, to go on doing it.

Arakawa:

Tremendous. And even if we can handle this expression, we don’t know such boundary conditions for this scale. If we dare use this scale, then we have to worry about the [???] [laughs]. So it’s a tremendous scale.

Edwards:

Yes.

Arakawa:

Now, then there were models based on the loads governing this one, and then do we know that those govern with this, including the interactions all of this and the basic distant load, this kind of thing, we have to say we don’t know. We know through exam and we know better and better, but that’s important. So in that sense, much effort of the modeling is to find those laws and is not the pure mathematical task. Okay, so now your question is about the observations. Okay, so to find these loads, how we know, well, find the load so it’s not mathematical, so we cannot just, say, do that on paper, we cannot just do that with a computer — we have known, we have to regard it as how it actually behaves, so observations are very important. But if I’m asked the direct in fact observations to actual, more there, then I have to say that link is not so bad. Observations are very important to know the nature and inspire the idea of how we formulate this complicated system. We cannot do that by just going out of here. Okay. For example, on formulating the effect of this, cumulus clouds, on this, the problem again is parameterization. So I am doing some of the observational work here, but…

Edwards:

We’re talking about the mezo scale of deep convection and shallow convection.

Arakawa:

Even some observation, one, or using the models for this. But it’s hard to see the [???] effect on this one. So in that way we are learning a lot, but eventually we’ve got to ask how to formulate it. So observation is a description. Then we have to go through that stage, but the model requires more than that. Then we have to formulate. That’s the way we can explain why nature is behaving that way, and do we have to predict it.

Edwards:

So, just in thinking back about this early period in the 1960s, did you do a lot of comparison of your models with what data existed?

Arakawa:

We know how the mean general circulation looked like. That is a tremendous advance in post-World War II, there was that period and the earlier, it’s behavior in World War II, that part of the observation that we improved a lot. So around 1950, we had a reasonably good part of the observation. That’s why Chinese developed von Neumann, and appeared had some success. So the initial condition was — of course taken from observation — was reasonably good, and verification was available after that.

Edwards:

Oh, this brings up another question. What you’re saying is true for short time scales. Obviously after World War II there were more upper air observations, but only for a few years, and we have an increasingly long record now of upper air observations, but that didn’t exist at this time, and the Charney-Fjörtoft-von Neumann model was for a day or two of general circulation. But let me ask you this before you go on: What was a typical length of a model run? How long did you simulate with each successive generation of models here?

Arakawa:

Longest run we had in 1969 or so was for three years. I think that was the longest at that time. And then we saw even with everything remaining the same that one year is different from another, you know, for entire year change. And then, of course, out of the broad experience that I had. But after that, maybe even after this period, really we knew how general circulation looked like, how there were maps that looked like in the mean. Before that we didn’t leave any space there, only for observation, and the model had to have the time to produce something similar to that. That around here (early ’70s) to improve the model, we have to understand the main processes, adjust the model, for cumulus parameterization, for example. We didn’t know much about this. Then there were what different field observations can provide. That is one of those is basically the motivation for observational system, but it was good around the time that we cannot easily observe. But later on became a very big advocate of that modeling component. So basically, all observation around people. So it’s good if, but with field experiments. And one of the most developed in terms of intensity through the experimental work by Gate. 1974. Motivation of the experiment, that his experiment.

Edwards:

Right, right.

Arakawa:

So that is very useful data, even now. And what the idea is to try to understand collective behavior of cumulus convection. And the improvement derivation. So, there might have been optimism if we had done such intense observation we understand the weather variable much better, then they would promote parameterization, maybe so. But in that sense it’s a failure.

Edwards:

Explain that, say that again. I didn’t quite get that.

Arakawa:

Okay. It’s a very intensive observation. And there’s a hole. If we do such an observation, we can understand that cumulus convection much better, so that the problem of parameterization may be solved.

Edwards:

Yeah. But then you said it was a failure.

Arakawa:

Failure of that optimism. Okay? We did find so many things. And there they found parameterization is much more predictive than if you. You.

Edwards:

Right, right.

Arakawa:

So it’s not a failure; it’s a success. But in theory it is a success.

Edwards:

But it wasn’t as simple as you hoped it might be.

Arakawa:

Right. So, in that sense observation, I cannot recall any example for me directly. Observation inspires idea. Observation doesn’t give formulation. Nature doesn’t have a [???].

Edwards:

Yeah, yeah. But it must affect how you parameterize and tune the models.

Arakawa:

No. The observations can help with tuning, but the tuning is just to fit the various selective situations that it really doesn’t try to pass that quite through the model. For example, if we could tune that, then the support will decrease on these sides as opposed to the mathematical model, and then the overall [???] model usually becomes worse, in part because we don’t know the [???] in this equation, so that we don’t know the equations, but that may be just a fact that varied for that particular scale. I’m thinking that we pick up new things, so the equation may have been modified. And also if we cure the results that is tuned for the results, our data will fit this observation. But if we decrease observation, then we already tuned. If we add a new physical thing, which is a definitely good thing to do because the nature has such new things, but usually the model results from us. So observations, then we get some of the chance into the model to recreate that. Well actually, tried to understand why that’s so and what process produced that. And then they were formulated. But this is a delicate issue. I advocate for the importance of collaboration, okay. But it still doesn’t automatically change the model. It’s up to us make use of the observation. All right? Just the existence of certain things in nature does not mean we can model them.

Edwards:

Were there any points in this history where observation had a particularly significant impact on what you did?

Arakawa:

To me personally?

Edwards:

Yes, and to the modeling group as a whole.

Arakawa:

In the modeling community, of course, observations that we got, that were included in the research programs, had tremendous impact from edification. And there were products and we knew more about nature certainly, so that we recognized more clearly what we had to do. And of course for weather prediction, not necessarily just a day but over ten days and so on, and the medium range forecasts, the observation is crucial to, but of course we have to treatment with those raw data, that’s programming data assimilation. So. And also, most focused observation will agree in general because it gives us some idea of the [???], but what don’t we have to do.

Edwards:

Right, right.

Arakawa:

Actually I’m using the Gate observation almost consistently. It is ten years old observation. Those data are used by many people, but I am still using data from the new way of looking at. In fact, it’s more where we can define new problems in the parameterization, then we look at the data from that point of view, and have now that data for analysis and we can try to do that. We use the data still, but available is new data, of course. But for my purpose, the more intensive observations, not necessarily global, but intensive more focused observations is more useful. Because GCM just synthesizes everything. There was very logical saying that one pair of eyes complemented by another pair. So we can get the right report by wrong means! Of course the same hasn’t been tried; there was no problem like that. But just to keep very fine, we said that not to give promise. Just very important to us is the current state. That doesn’t really guide in the future. In the future we have to have more focus, not having everything together. I think there is a figure similar to that.

Edwards:

Yes, it is. This is the chart containing the cumulus parameterization model.

Arakawa:

This one includes everything. So if the result is wrong, it can be very wrong due to this or due to this, or it is wrong here with this boundary here, but actually the cause may be this one or this one. So to me, the only way I can get confidence about the model is to make sure that each piece is right. And there were many works outside of that, you see. They improved the treatment of clouds, use a cloud model.

Edwards:

Let’s talk about that a little bit, because obviously this has been very important to you. You’ve written about it throughout your career, cumulus parameterization issues and clouds in models. One of the arguments that comes up again and again in the political debate about climate change is the models don’t represent clouds well enough. This is sort of Richard Linsin’s [?] argument. The problems at that level are so large that we can’t trust the models at all. What’s your feeling about that? What’s the state of the art in cumulus parameterization now?

Arakawa:

Well, it’s more than the problem of cumulus parameterization.

Edwards:

Of course, or cloud parameterization in general.

Arakawa:

Clouds can be parameterized only when cumulus can be parameterized. Cumulus can deposit the moisture to upper layer, and upper level clouds are important for global warming and such things. I can certainly understand what Linsin says. Not necessarily I agree with him completely. In fact I don’t necessarily agree with every detail of his argument. But I kind of agree with his overall assessment. He goes extreme in a way [chuckles]. But I certainly agree with him, there’s a great answer that is involved. I know that the model is far from nature. But there are differences in Linsin and myself. Linsin thing is anti-GCM, almost as. I’m directly working on GCM, but I have an observation of what GCM can do. Linsin is saying GCM, on the other hand, is [???], so that’s an important difference. He and I more or less are similar on some aspects, but with a main difference. I often agree with him, and basically the future GCM would solve the problem with the GCM. And what I said might have taken as kind of a. In fact, somebody asked, oh boy, that you have done lots of things with that, but you are not recognized for that. Of course, understanding we have done a lot [chuckles]. Not me, I mean in general. But still, there is a long way to go. But we have done a lot in the sense we know the problem we are facing is very challenging…

Edwards:

Challenge problems.

Arakawa:

Yes. Before, well, I said magnificent the way the GCM expanded in studying form that were there, Phillips had the model that really doesn’t involve the formulation of heating; heating is prescribed. I kept expanding the other one, like that, and then you have the hydrology, like that.

Edwards:

Why did you change the word “magnificent” for that? I thought that was very interesting.

Arakawa:

Oh, this happened to be… [Chuckles] Since English is not my first language, not sure that word “magnificent” is appropriate or not.

Edwards:

I like it a lot; I’m just curious what it meant to you.

Arakawa:

This is basically a version (charts from the article of the history of modeling). This is first phase (EPOC [?] making phase). This is by Charney, and highlight is Phillips. That’s [???] making. But only this.

Edwards:

So it’s only dynamical processes. The second phase we get a hydrological boundary layer, radiation clouds, precipitation, and land temperatures are described.

Arakawa:

I think that this third expanding is very magnificent! [Laughter]

Edwards:

I agree!

Arakawa:

All this is just a small portion of what the Meteorology Committee was interested in, like that. But now it tries to cover all areas of meteorology. It is magnificent — a spectacular expansion of the process complication in the model, spectacular expansion by the people who are getting bored, and the amount of computation in many ways was I think magnificent. Really! [Chuckles] Like the construction of…

Arakawa:

…in that period the SSD is fixed. Well, there are exceptions I measured with Manabe, and it did work with a couple, his model. But I used more the NW work on the GCM, using this period as fixed. That means we get very, very important part of the answer from the climate point of view. Climate is so sensitive to SSD, if SSD is fixed, then the atmospheric temperature near the surface cannot be far away from this because it is strongly coupled. Then above that the temperature gradient is more or less determined in the topics by GMS convection. Then the temperature of the inter-troposphere cannot be so bad, even when actual process is formulated completely wrong. But all of that, the apparent advantage just disappears when we coupled atmospheric model with ocean model. Then there’s climate drift — climate goes into a very different climate. So one standard technique that we use is the flux adjustment. Now, Manabe had used this, and Manabe tries to defend it. Some people really criticize. I’m sort of in the middle. For Manabe’s purposes that’s fine, but not everybody has the same purposes as Manabe. Because Manabe’s real concern with what would happen say [???] rate out, the efficacy would do with that, based on the best knowledge we now have. But even the best knowledge we now have, there is climate drift or like that, that adjustment. Things like that I have no problem; I think that’s very useful. But that’s not the whole problem because that can hide what’s wrong in the model. So here we don’t use flux adjustment.

Edwards:

Never have?

Arakawa:

Never. If something is wrong with the model, I would like to see it. Of course not everybody should be like me. If everybody is like me, then there is no progress!

Edwards:

Something you said a minute ago brings up another issue I wanted to talk about, which is to what degree have you studied the carbon dioxide issue in your work?

Arakawa:

Not directly. At one time, late 1970s, on that occasion I got involved with that, the National Academy of Sciences formed a panel to assist the work on the side effect, and the chairman of that panel was Jule Charney. Charney called me up to join the panel. I told him I’d never worked on that program. Then the Charney said, “That’s exactly why I want you! So it can be objective.” So I served on that panel with a few others and we met a couple of times, and Charney also came here. And they produced a report for the National Academy of Sciences. There was a similar one, a later one, but what I’m talking about is the panel report chaired by Charney late 1970s. My role there was to give a sort of assessment of the GCM experiment.

Edwards:

And what did you think of it, at that time? What did you have to say?

Arakawa:

Actually not only me, but the panel, well, the others were already [???], but nobody was directly involved. The CO₂[?], but all of them are expert of course of a related field, but not…

Edwards:

So Manabe was not on that panel?

Arakawa:

…no. Manabe was invited to brief us. Jim Hansen and so on gave their input, but we were supposed to have the panel’s own judgment. So we kind of tried to stop from some skepticism and tried to find something wrong. The big possibility is the missing negative feedback. Nobody argues against the direct effect. That direct effect of carbon dioxide is not so large, but there is positive feedback, quiet feedback, there’s no feedback, and so on. Leading questions at one of the meetings was water vapor feedback. The panel tried to look for the missing negative field feedback. If there is negative feedback, the prediction has negative feedback missing in the GCM at that time for that kind of research, then the results may be wrong. The panel conclusion was that we were not able to find the missing negative feedback, so we cannot say that the results are right, but we didn’t find anything wrong.

Edwards:

So no major reasons to question that, but you didn’t conclude that they were perfect.

Arakawa:

So eventually we basically endorsed the results, and that goes to Charney’s attitude.

Edwards:

Why? What do you mean by the Charney attitude?

Arakawa:

That that is the panel’s conclusion. Of course how Charney feels is of course dominated. Charney was a general [?] with the panel. I said it is an important and interesting problem, but I didn’t do that problem in the past and probably not in the future because, well, it’s hard to understand the interaction that this process is keeping me to, busy enough. But I’m not saying everybody should do the same thing.

Edwards:

This takes me back to something we were talking about before, which is the length of your model runs. You said in the 1960s the longest run was about three years. How long do you run the models for now, and how has that evolved?

Arakawa:

The group here is sort of model development oriented. Usually to see the impact of some of the regions, it doesn’t need a very long run. So this group is certainly not representative, but we did have long runs. For atmosphere longest run, I don't know, some tens of years, maybe fifty years, something like that. And the top model, some tens of years. Manabe must have been using his model over thousands of years.

Edwards:

Another thing. In the last decade or so there has been a lot of interest in aerosols. Has that been part of your work? Is that something you have included in the models?

Arakawa:

I have never worked on that, although I recognize sometimes why it should be included. But now we don’t have that manpower. Some of the other programs, they do.

Edwards:

Let’s go change to a somewhat different but related issue, which is the political context of all of this. Obviously you are primarily a theoretical modeler. I haven’t been involved in the climate change controversy too much. But do you recall any significant influences of the political climate on your work? For example, when the SST issue, the supersonic transport issue came up in the late ’60s, early ’70s, we talked yesterday about the CLIMAP study that the Department of Transportation sponsored later, and you got some money from them for your work. Have there been other episodes like that, where you saw some connection between what you were doing and this context, especially through funding?

Arakawa:

The National Science Foundation is consistent sponsor in the research.

Edwards:

By the way, have there been any particular grants officers at the National Science Foundation who have been connected with you for a long time?

Arakawa:

This is over a very long time, so the officer there changes. But for the last 15 or 20 years this grant is administrated by Pamela Stephens. The name of that program is NASSCO Dynamics, program of NASSCO Dynamics. The branch that can support us is the Climate Dynamics program. In a certain period, different programs jointly supported, but Pamela says that is up to us and you don’t have to decide which. So Pamela consistently… Before that, much earlier, Gene Bierly, who was the head of the Atmospheric Science Section until a few years ago, he was a very early supporter of the project here on even a personal level. There were no barriers.

Edwards:

Anybody else you think of?

Arakawa:

At NSF before Pamela Stephens… Well I think Pamela is long enough. NASA is a big supporter.

Edwards:

Since the early ’70s?

Arakawa:

Right. And Milton Halem was a big supporter, but he moved to another branch several years ago.

Edwards:

And then the Department of Transportation and the CLIMAP…?

Arakawa:

Was also early 1970s. But recently the Department of Energy and the global changes group.

Edwards:

Any people in particular you think of there?

Arakawa:

Department of Energy has similar programs, and the program we get the biggest support from is called CHAMP. This is a relatively recent one, a joint project with Majoso. He’s nearly retired, but he is mainly handling the administrative aspects. I think the person at the DOE in charge of this is Bader. Majoso is handling.

Edwards:

Back to this question of political contexts, can you think of any particular episodes where you felt that that was an important factor in what you were doing, or your ability to get money from these agencies?

Arakawa:

I’m not a political man; I want to stay away from it when, of course, I need support. Of course there are fluctuations, but fortunately in the past there were not really serious problems. I guess the GCM is [???] purpose. When their medium-range, long-range predictions are precise, of course it has to depend on GCM. Carbon dioxide problem and global changes in precise, that depends on GCM [???]. It’s really not the purpose. So objects of what is involved are basically… it’s relatively free from the political. We are working on the say carbon dioxide problem, I think I’ll get more influence. So that support is pretty persistent, but NSF and NASA have consistently supported us. But other agencies change. The Department of Transportation in the early 1970s, then later the US Navy.

Edwards:

Oh yeah? Who was that?

Arakawa:

Thomas Rosmond was very interested in medium range prediction, and they were interested in our model, so they supported us. That was a very complicated process. Rather than they give us money directly, they gave money to NSF so it may supplement the NSF grant to us. That continued for a few years. But the changing money from one agency to another took a long time and was very complicated. Now this is the Department of Energy that was one of the big supporters, so that reflects political situations. [Chuckles] But for us, what we were doing, it has all been consistent.

Edwards:

What about Jule Mintz, did he have involvements in political issues that you’re aware of?

Arakawa:

He was a very active person in many, many ways. I don't know that he can be called political. He is a very aggressive person. From the middle of 1970s he regularly took sabbatical and went to Israel at least for one quarter, and then he was on sabbatical leave of absence, and he finally resigned, when, in the early 1980s and moved to NASA. But NASA, there’s Milton Halem. I think Mintz was relatively free from the politics in NASA, if there is any. So I think he enjoyed just doing research there. So I cannot answer your question. He is a very outgoing, aggressive person, but I hesitate to say he is political.

Edwards:

Did he serve on a lot of panels, like the one you mentioned of the NAS. That seems like something he would likely be invited to do. Do you remember any of those that may have been important?

Arakawa:

He had contact with a very broad variety of people, but I don’t characterize his activity as political.

Edwards:

A couple more questions about issues that might were politicized, but I’m just interested in your general involvement with these, if any. When was the nuclear winter issue? Because your colleague here, Richard Turco, was involved with that with the Turco-Sagan group that came up with that idea in the early 1980s. Did you participate in that debate at all? Did you have discussion with him? Because one of the criticisms that is often leveled against the Turco-Sagan group is they used a very simple climate model to get their projection of a nuclear winter, and when the same kinds of forcings were applied to GCMs it didn’t show nearly the same level of effect.

Arakawa:

That kind of work will be criticized anyway. I may be the one who knows the GCM very well, but that includes its weakness. So I don’t object to that research; in fact I’m interested in seeing that. But that’s something I am not interested in myself because I am aware of so many problems. But I don’t say what other people are doing — I mean that’s very interesting. That represents our current state of art, so that’s fine. A similar thing can be said to nuclear winter, maybe to a different degree. And of course real climate change cannot be said without involving the ocean, and involving the ocean raises some bad problems. But I think what really matters in the constant degree, the tendency to know that kind of thing to happen is pretty [???], but I don't know how correct those numbers are. Yeah, it’s very easy to criticize. In that sense, even the most advanced GCM right now can be criticized. It’s a state of are from nature. And most GCMs have the difficulty when coupled with the ocean if the flux is not adjusted that it shows a very serious weakness. Flux adjustment may be a good way to stay away from that weakness for the time being, but of course that’s not really the objective of the model to stay away from the problem, not solving the problem. So although he is a colleague here and he is a very good friend, I really don’t know the details of that work. Of course I know about it, but he did that work before he came here. Actually impact will be very different if we consider the ocean. But involving the ocean can delay the time using carbon dioxide with the ocean, depending on how deep the ocean that is involved, it can delay. Warming may take place eventually, just answered in this way. Now in the nuclear winter case it is a bit more subtle. It is a very much shorter time scale. Then the ocean delays significant, then the overall effect may change significantly, quantitatively speaking. So that may be answered in this more. But still… [chuckles]

Edwards:

Another issue is the ozone depletion issue, and that you said yesterday became part of your model, you put ozone chemistry into the model in the CLIMAP study.

Arakawa:

But it is not doing a very good job. That is very simple, and we put in that to satisfy DOT’s interest in the early 1970s, we put in that. We have been using that, but that is so poor and too simplified that it’s not doing good for the model; it is rather doing harm. Because of the oversimplification, the ozone prediction is not very good, and that hurts the model. So if we just prescribe ozone, giving up prediction of ozone and prescribe the ozone, then the model results are better. But of course versions that we are using, that isn’t prescribing the ozone because we had our purpose as something else, at least right now. But that report covered with Turco and Majoso, I think that Turco and his people are doing a lot. There are many traces of Turco’s stuff in the GCM.

Edwards:

The models you’ve done haven’t prescribed ozone.

Arakawa:

No, his people are using our GCM as well as other models. A man working for him is working with the GCM and putting traces on the chemistry, and also the ocean. That is a big project of Majoso, Turco and myself. I think that is sponsored by NASA. It is a very big project. If you are interested in very recent kind of status, it may be well for you to see Majoso because I’m trying to phase out from administrative aspects.

Edwards:

The problem is I’m interested in so much! For this project I want to focus on the historical things. But Mojoso sounds like he has been important for a long time in other areas, too.

Arakawa:

From 1980 as far as GCM is concerned, middle of 1980s.

Edwards:

I have one last question, and this is sort of grab bag question, and it’s because this invention will become part of the historical record of your work. So I wanted to ask you what, in your own view, are the most important things that you have contributed to climate science? Obviously there is an awful lot, but if you had to pick three or four things that you think are your very most important contributions, what would you say?

Arakawa:

It’s hard to say. In terms of scientific literature is concerned, I mentioned the 1966 paper, so-called the Jacobian paper. It deals with computational aspects but nonlinear, but really not really mathematics [?] approach rather than from the point of view of the physics of discrete system. That is has become a famous paper, and it is very broadly cited, not only meteorology — just check the citation index and journals I never heard of cited that paper in engineering and wherever fluid is involved. That is a very broad citation. And number of citations still remains constant in time. Number of citations is constant in the Arakawa-Schubert paper of 1974. As I told you, that is not a very difficult paper to understand, or I didn’t think so. Even before that was published I gave an oral presentation that became already kind of famous. Well, it is famous, but that does not mean all people agreed. There are two types of reactions to that paper: one is from the modelers and theoreticians that say it’s too complicated. There are criticisms the observation of cumulus clouds is too simple — clouds are not that simple. Then of course this paper might have tried to make some simple things unnecessarily complicated. That’s a very negative way of looking at it. Of course I didn’t look at it that way, but I’m sure some people looked at it that way. Relatively recently that’s changing partly because more people recognize the problem is complicated and began to appreciate the depth of this paper. Before, following this approach was out of the question for operational models — it’s too expensive. But the computer became faster, and also some people worked on this to make the model more efficient without losing essence. There are many names: Durock, Arakawa, — there are many names. Some of those are using the operational models. The number of operational models to use is increasing recently, so it’s on reason the citation of this paper is even increasing a few years after, so maybe I published that paper too early. Another paper, you have Arakawa-Lamb, which is basically a description of the UCLA GCM published in 1977. But inside there were many things new at that time. Some people call that the bible of modeling. So that paper is still being cited. Especially the A, B, C, D, E grid. The choice of that grid influences adjustment mechanisms, and also influences the existence of the computational model. There are other papers, but those three are the most frequently cited. But I guess my overall contribution to the society is not just through individual papers but this collectively. The model here had a broad influence at institutions, and I think is my main contribution.

Edwards:

Well, great. Thank you so much. This has been incredibly interesting.

Arakawa:

That’s right, yes. I integrated that to my lectures and the paper presented at that Symposium. It was my lecture at the Korean Meteorological Society where I further expanded adjustment and instability mechanisms in the atmosphere and atmospheric models. Okay, so that includes the computational instability, but of course it reviews the physical instabilities. So now your question is how I…?

Edwards:

What’s the relationship between them? And I’m especially interested in how important computational instability is relative to physical instabilities and how that’s changed over time.

Arakawa:

Computational stability is a necessary condition. There are more than one way of defining stability, but usually the most simple-minded definition of stability is that the model runs without blowing up. Okay? And that’s a necessary condition. But usually it is far from sufficient conditions — model may keep running, but producing garbage. That is still stable in view of the definition of stability pretty commonly used. So what I have been advocating is computational instability is of course too bad, but that is not enough to think. What we do with discrete system, we discretize finite difference, spectral, that really doesn’t matter. Since the original system is continuous system, in that way we can tremendously distort the behavior of solution. And instability is just an example, one bad example. But even a stable solution can be very different. In a discrete case, even there is a mode which doesn’t have a counterpart in the continuous system. That is computational mode. And so one thing I am advocating is to maintain the analogy of the discrete system to the continuous system, and the computational stability problem is just a small portion of it. My 1966 paper, the motivation of that paper was to overcome the practical problem of nonlinear computational instability, a relatively newly found computational instability at that time, and so it is nonlinear so it’s difficult to get the mathematical analysis of the nature of the instability. Linear instability is relatively easy. The nonlinear instability was a new kind of instability discovered originally by the Phillips experiment in 1956, because his model blew up. The usual computational instability is eliminated or reduced by making ΔT short. That of course he did, but it didn’t help; it can only just slightly delay the catastrophe. So that there is a new kind of computational instability. Phillips a few years later called it the nonlinear computational instability. In my 1966 paper, the motivation is to overcome nonlinear instability so that we can construct a GCM. But the other point, that paper, it’s more than that. It’s really not a mathematician’s traditional approach. Really I think I’m a kind of physicist, physicist approach, just this physical behavioral of discrete system should be analogous to that of the original system. So that is the basic, but in addition to that there are real physical instabilities. Of course in nature, the instability can exist, but if the disturbance grows due to physical instability, they were eventually equilibrated. That’s why nature doesn’t blow up. Or instability may exist, but may not be actually realized, depending on the scale or the other conditions, that is, potential for instability may exist but it may not be realized. So my overall idea is not really just try to have some discrete system — we have to have a discrete system which is physically analogous to the continuous system.

Edwards:

Right. To what degree do you feel that the… I guess I’d want to say computational instabilities in the models have been reduced to insignificance at this point. Are there still major areas, major issues, for example truncation errors and things of that sort?

Arakawa:

The idea with the truncation error, it’s a bit old idea, I shouldn’t say unimportant, but that doesn’t tell the whole story at all. And what I think very important is the computational mode. It’s a mode which does not have a counterpart to the original equation. So it’s spurious so that we cannot talk about the error. The error is the difference of that solution from the true solution, but there is no true solution corresponding to that. So for that kind of mode, the concept of truncation error is useless. So now that computational mode usually appears with very small spatial structure, either horizontally or vertically, or it can be in time actually. That thing was first found in time, but now it is in space, it can be a space in the vertical. And that appears with the small scale. But when the model doesn’t have much physics, especially when model doesn’t have moisture, it may not matter so much and the model result may have a noise we can smooth just cosmetically. But if there is a moisture, then condensation, precipitation may appear after responding to that small scale. And whenever condensation is involved, it is subject to rectification. Suppose… [Draws a wave form on a piece of paper] Usually condensation appears when the vertical velocity is upward. Suppose vertical velocities are like that. The abscissa is time or it can be horizontal.

Edwards:

Up and down, according to the wave. Changing vertical velocity, positive and negative.

Arakawa:

Right. Then the condensation appears when the vertical velocity is upward, but nothing happens here.

Edwards:

I see. All right, so we get a series of sort of lumps of condensation with no condensation on the downward side of the wave.

Arakawa:

Another example of rectification is from AC to DC. Now, each of these may have the shorter time scale or the shorter space scale. We may be interested in a much larger scale, but if this kind of situation happens, this just adds up. So this is another way of saying that the cumulus parameterization is important. This may be the size of cumulus, but that adds up. Now, suppose this kind of situation takes place and is associated with computational mode, which is a kind of zigzag structure. Then that’s horrible. It adds up. So it’s a nonlinear interaction. But usual nonlinearity we are aware, in fluid dynamics in general, is due to advection, velocity times the gradient of that variable. That’s nonlinear. Nonlinearity we are familiar with. That’s in the form of product—quadratic. But this is not that kind of nonlinearity; this is not quadratic, more on and off type, recognized sign. In my mind, I don’t distinguish too much of the computational from the physical. I’ll just say I’m concerned with the physics of the discrete system — it can be computationally distorted. But pure computational instability is just a very small portion of the problem.

Edwards:

Well, this sort of leads naturally into another question, which is a whole area we haven’t talked about at all yet, which is validation of your models. When we were talking yesterday in the article you wrote about Yale Mintz, you talked about how important observation was to him, and your early background was also in observational meteorology. So clearly both of you were strongly aware of the real atmosphere, and you have just been talking about how difficult it is to model in a discrete form, so I’m curious about to what degree and when you began to compare the model results and how with what sorts of data. Part of what I’m interested in here is that during this whole period of your modeling efforts, the amounts of data that became available were increasingly large, a very rapid, upward curve, and especially in the seventies, when satellite data became available. So I just want to hear about how that influenced your work, if it did.

Arakawa:

I have to be a bit careful [laughs]. Of course we’ve got to know how nature behaves to model nature intelligently, of course, so modeling has not become a pure mathematical task. Really, and I think I quoted at the opening in that short lecture, and also Rossby had the same feeling, that we don’t know the basic equations. Okay. Many people, meteorologists, don’t recognize that. [Static on recording begins] Okay, we know the laws involving molecular viscosity and molecular heat conduction, and those processes are only for the scale of millimeters.

Edwards:

Right. Molecular scale processes.

Arakawa:

Right. [Retrieves chart]. This is the horizontal scale. And in this one we are interested in this scale.

Edwards:

The planetary and synoptic scale.

Arakawa:

They are in this spectral range, 100 to 10,000 kilometers. And we know them pretty well. But…

Edwards:

The one millimeter to one centimeter, which is on this chart also.

Arakawa:

But that is not all.

Edwards:

Right, to go on doing it.

Arakawa:

And even if we can handle expression of these small scales, we don’t know boundary conditions for those scales. If we dare use those scales, then we have to worry about the butterflies [laughs]. So it’s a tremendous scale.

Edwards:

Yes.

Arakawa:

Now, then there were models based on the laws governing large scales, and then do we know laws that govern those scales, including the interactions between all of them? This kind of thing, we have to say we don’t know. We know through experience and we know better and better, but that’s important. So in that sense, much effort of the modeling is to find those laws and it is not the pure mathematical task. Okay, so now your question is about the observations. Okay, so to find these laws we cannot just, say, do that on paper, we cannot just do that with a computer — we have to know how it actually behaves, so observations are very important. But if I’m asked the direct impacts of observations on actual modeling, then I have to say that link is not so strong. Observations are very important to know the nature and inspire the idea of how we formulate this complicated system. We cannot do that by just going out of here. Okay. For example, on formulating the effect of this, cumulus clouds, on this, the problem again is parameterization.

Edwards:

We’re talking about the scale of deep convection and shallow convection.

Arakawa:

But it’s hard to see their net effect on this one. So through observations we are learning a lot, but eventually we’ve got to ask how to formulate it. So observation is a description. We have to go through that stage, but the model requires more than that. Then we have to formulate. That’s the way we can explain why nature is behaving that way, and we have to predict it.

Edwards:

So, just in thinking back about this early period in the 1960s, did you do a lot of comparison of your models with what data existed?

Arakawa:

We know how the mean general circulation looks like. That is a tremendous advance in post-World War II. So around 1950, we had a reasonably good part of the observation. That’s why Charney developed the first NWP model with von Neumann, and it appeared that it had some success. So the initial condition was — of course taken from observation — was reasonably good, and verification was available after that.

Edwards:

Oh, this brings up another question. What you’re saying is true for short time scales. Obviously after World War II there were more upper air observations, but only for a few years, and we have an increasingly long record now of upper air observations, but that didn’t exist at this time, and the Charney-FjØrtoft-von Neumann model was for a day or two of general circulation. But let me ask you this before you go on: What was a typical length of a model run? How long did you simulate with each successive generation of models here?

Arakawa:

Longest run we had in 1969 or so was for three years. I think that was the longest at that time. And then we saw even with everything remaining the same that one year is different from another, you know, inter-annual change. In any event, we knew how general circulation looked like. The model had to have time to produce something similar to that. Around here (early ’70s) to improve the model, we have to understand the main processes, cumulus parameterization, for example. We didn’t know much about this process. That kind of knowledge is what different field observations can provide. But around the time we cannot easily observe such things. But later on Carney became a very big advocate of field experiments. And one of the most developed field experiments was the GATE 1974

Edwards:

Right, right.

Arakawa:

So that is very useful data, even now. And the idea was to try to understand collective behavior of cumulus convection. And the improvement on its formulation. So, there might have been optimism if we had done such intense observation we understand the weather much better, then they would promote parameterization, maybe so. But in that sense it’s a failure.

Edwards:

Explain that, say that again. I didn’t quite get that.

Arakawa:

Okay. It’s a very intensive observation. If we do such an observation, we can understand that cumulus convection much better, so that one may think the problem of parameterization may be solved.

Edwards:

Yeah. But then you said it was a failure.

Arakawa:

Failure of that optimism. Okay? We did find so many things. And we found parameterization is much more difficult than if you might have thought before the experiment.

Edwards:

Right, right.

Arakawa:

So the experiment itself is not a failure; it’s a success. In theory it is a success.

Edwards:

But it wasn’t as simple as you hoped it might be.

Arakawa:

Right. So, in that sense. I cannot recall any example of observation that improved parameterizations directly. Observation inspires idea. Observation doesn’t give formulation. Nature doesn’t have a parameterization.

Edwards:

Yeah, yeah. But it must affect how you parameterize and tune the models.

Arakawa:

No. Observations can help tuning, but the tuning is just for selected situations. It really doesn’t pass that knowledge to the model. Well actually, we must try to understand why that’s so and what process produced that. And then they must be formulated. But this is a delicate issue. I advocate for the importance of observation, okay. But it still doesn’t automatically change the model. It’s up to us to make use of the observation. All right? Just the existence of certain things in nature does not mean we can model them.

Edwards:

Were there any points in this history where observation had a particularly significant impact on what you did?

Arakawa:

To me personally?

Edwards:

Yes, and to the modeling group as a whole.

Arakawa:

In the modeling community, of course, observations included in the research programs had tremendous impacts. And there were products through which we knew more about nature certainly, so that we recognized more clearly what we had to do. And of course for weather prediction, not necessarily just a day but over ten days and so on, and the medium range forecasts, the observation is crucial to initial conditions, but of course we have to do adjustment with those raw data, that’s program called data assimilation. And also, most focused observation will give us some idea on what we have to do and what we don’t have to do.

Edwards:

Right, right.

Arakawa:

Actually I’m using the GATE observation almost consistently. It is more than twenty years old observation. Those data are used by many people, and I am still using the data from the new way of looking at. In fact, it’s where we can define new problems in the parameterization, then we look at the data from the new point of view. We use the data still although new data became available, of course. But for my purpose, the more intensive observations, not necessarily global, but intensive more focused observations is more useful. Because GCM just synthesizes everything. In the future we have to have more focus, not having everything together. I think there is a figure similar to that.

Edwards:

Yes, it is. This is the chart containing the cumulus parameterization model.

Arakawa:

This chart includes everything. So if the result is wrong, it can be very wrong due to this or due to this, but actually the cause may be this one or this one. So to me, the only way I can get confidence about the model is to make sure that each piece is right. And there are many people working outside of that, you see. They improved the treatment of clouds, using a cloud model.

Edwards:

Let’s talk about that a little bit, because obviously this has been very important to you. You’ve written about it throughout your career, cumulus parameterization issues and clouds in models. One of the arguments that comes up again and again in the political debate about climate change is the models don’t represent clouds well enough. This is sort of Richard Lindzen’s argument. The problems at that level are so large that we can’t trust the models at all. What’s your feeling about that? What’s the state of the art in cumulus parameterization now?

Arakawa:

Well, it’s more than the problem of cumulus parameterization.

Edwards:

Of course, or cloud parameterization in general.

Arakawa:

Clouds can be parameterized only when cumulus can be parameterized. Cumulus can deposit the moisture to upper layer, and upper level clouds are important for global warming and such things. I can certainly understand what Lindzen says. But not necessarily I agree with him completely. He goes extreme in a way [chuckles]. I know that the model is far from perfect. But there are differences between Lindzen and myself. Philosophically Lindzen is anti-GCM while I am not, of course. I’m directly working on GCM and I have feeling of what GCM can do. But still, there is a long way to go. But we have done a lot at least in the sense we know the problems we are facing is very challenging. Staying away from GCMs does not solve the problems.

Edwards:

Challenge problems.

Arakawa:

Yes. Before, well, I said in a magnificent way the GCM expanded. Phillips had the model that really doesn’t involve the formulation of heating; heating is prescribed. We kept expanding the other side, like that, and then you have the hydrology, like that.

Edwards:

Why did you change the word “magnificent” for that? I thought that was very interesting.

Arakawa:

Oh, this happened to be… [Chuckles] Since English is not my first language, I was not sure that word “magnificent” is appropriate or not.

Edwards:

I like it a lot; I’m just curious what it meant to you.

Arakawa:

This is basically a version (charts from the article of the history of modeling) used in the first phase (epoch-making phase). This is by Charney, and the highlight is Phillips. That’s epoch-making. But only dynamics.

Edwards:

So it’s only dynamical processes. The second phase we get a hydrology, boundary layer, radiation, clouds, precipitation, and land temperatures are described.

Arakawa:

I think that this expanding is very magnificent! [Laughter]

Edwards:

I agree!

Arakawa:

The first phase is just a small portion of what the Meteorology Community was interested in, like that. But now it tries to cover all areas of meteorology. It is magnificent — a spectacular expansion of the process complication in the model, spectacular expansion by the people who are involved, and the amount of computation in many ways was I think magnificent. Really! [Chuckles] Like the construction of…

Arakawa:

…in this period the sea surface temperature (SST) is fixed. Well, there are exceptions: Manabe did work with a coupled model. Fixing SST means we prescribe very, very important part of the answer from the climate point of view. Climate is so sensitive to SST, if SST is fixed, then the atmospheric temperature near the surface cannot be far away from this because it is strongly coupled. Then above that the vertical temperature gradient is more or less determined in the tropics by moist convection. Then the temperature of the troposphere cannot be so bad, even when actual process is formulated completely wrong. But all of that, the apparent advantage just disappears when we couple atmospheric model with ocean model. Then there’s climate drift — climate goes into a very different climate. So one standard technique that we use is the flux adjustment. Now, Manabe had used this, and Manabe tries to defend it. Some people really criticize. I’m sort of in the middle. For Manabe’s purposes that’s fine, but not everybody has the same purposes as Manabe. Because Manabe’s real concern is what would happen say next century based on the best knowledge we now have. Things like that I have no problem; I think that’s very useful. But that’s not the whole problem because that can hide what’s wrong in the model. So here we don’t use flux adjustment.

Edwards:

Never have?

Arakawa:

Never. If something is wrong with the model, I would like to see it. Of course not everybody should be like me. If everybody is like me, then there is no progress!

Edwards:

Something you said a minute ago brings up another issue I wanted to talk about, which is to what degree have you studied the carbon dioxide issue in your work?

Arakawa:

Not directly. At one time, late 1970s, on that occasion I got involved with that, the National Academy of Sciences formed a panel to assess the work on the CO₂effect, and the chairman of that panel was Jule Charney. Charney called me up to join the panel. I told him I’d never worked on that program. Then the Charney said, “That’s exactly why I want you! So you can be objective.” So I served on that panel with a few others and we met a couple of times, and Charney also came here after the panel meetings. And they produced a report to the National Academy of Sciences. There was a similar one, a later one, but what I’m talking about is the panel report chaired by Charney late 1970s. My role there was to give a sort of assessment of the GCM experiments.

Edwards:

And what did you think of it, at that time? What did you have to say?

Arakawa:

Actually not only me, but the panel. Well, nobody was directly involved in the CO₂problem, but of course all of them are experts of related fields, but not…

Edwards:

So Manabe was not on that panel?

Arakawa:

No. Manabe was invited to brief us. Jim Hansen and so on gave their input, but we were supposed to have the panel’s own judgment. So we kind of tried to start from some skepticism and tried to find something wrong. The big possibility is missing negative feedback. Nobody argues against the direct effect. That direct effect of carbon dioxide is not so large, but there can be positive feedback, negative feedback, no feedback, and so on. Leading questions at one of the meetings was water vapor feedback. The panel tried to look for the missing negative feedback. If there is negative feedback missing in the GCM used for that kind of research, then the results may be wrong. The panel conclusion was that we were not able to find missing negative feedback, so although we cannot say that the results are right, but we didn’t find anything wrong.

Edwards:

So no major reasons to question that, but you didn’t conclude that they were perfect.

Arakawa:

So eventually we basically endorsed the results, and that goes to Charney’s attitude.

Edwards:

Why? What do you mean by the Charney attitude?

Arakawa:

That is the panel’s conclusion. Of course how Charney feels dominated. Charney really lead the panel. It is an important and interesting problem, but I didn’t do that problem in the past and probably not in the future because, well, understanding the interactions between these processes is keeping me busy enough. But I’m not saying everybody should do the same thing.

Edwards:

This takes me back to something we were talking about before, which is the length of your model runs. You said in the 1960s the longest run was about three years. How long do you run the models for now, and how has that evolved?

Arakawa:

The group here is sort of model development oriented. Usually to see the impact of revisions, it doesn’t need a very long run. So this group is certainly not representative, but we did have some long runs. For atmosphere longest run, I don't know, some tens of years, maybe fifty years, something like that. Manabe must have been using his model over thousands of years.

Edwards:

Another thing. In the last decade or so there has been a lot of interest in aerosols. Has that been part of your work? Is that something you have included in the models?

Arakawa:

I have never worked on that, although I recognize sometimes it should be included. But now we don’t have that manpower. Some of the other programs, they do.

Edwards:

Let’s go change to a somewhat different but related issue, which is the political context of all of this. Obviously you are primarily a theoretical modeler. I haven’t been involved in the climate change controversy too much. But do you recall any significant influences of the political climate on your work? For example, when the supersonic transport issue came up in the late ’60s, early ’70s, we talked yesterday about the CIMAP study that the Department of Transportation sponsored later, and you got some money from them for your work. Have there been other episodes like that, where you saw some connection between what you were doing and this context, especially through funding?

Arakawa:

The National Science Foundation is a consistent sponsor of our research.

Edwards:

By the way, have there been any particular grants officers at the National Science Foundation who have been connected with you for a long time?

Arakawa:

This is over a very long time, so the officer there changes. But for the last 15 or 20 years this grant is administrated by Pamela Stephens. The name of that program was Large-Scale Dynamics. The branch that supports us now is the Climate Dynamics program. In a certain period, different programs jointly supported, but Pamela says that is up to them and we don’t have to decide which. So Pamela consistently. Before that, much earlier, Gene Bierly, who was the head of the Atmospheric Science Section until a few years ago, he was a very early supporter of the project here on even at a personal level. There were no barriers.

Edwards:

Anybody else you think of?

Arakawa:

At NSF before Pamela Stephens… Well I think Pamela is long enough. NASA is also a big supporter.

Edwards:

Since the early ’70s?

Arakawa:

Right. And Milton Halem was a big supporter, but he moved to another branch several years ago.

Edwards:

And then the Department of Transportation and the CIMAP…?

Arakawa:

Was also early 1970s. But recently the Department of Energy and the global changes group.

Edwards:

Any people in particular you think of there?

Arakawa:

Department of Energy has similar programs, and the program we get the biggest support from is called CHAMMP. This is a relatively recent one, a joint project with Mechoso. I think the person at the DOE in charge of this is Dave Bader.

Edwards:

Back to this question of political contexts, can you think of any particular episodes where you felt that that was an important factor in what you were doing, or your ability to get money from these agencies?

Arakawa:

I’m not a political man; I want to stay away from it except when, of course, I need support. Of course there are fluctuations, but fortunately in the past there were not really serious problems. I guess the GCM has a broad purpose. How precise medium-range, long-range predictions are, depends on GCM. How precise Carbon dioxide problem and global changes are depends on GCM. So the objective of what we are doing is basically…it’s relatively free from the politics. So that support is pretty persistent, NSF and NASA have consistently supported us. But other agencies change. The Department of Transportation in the early 1970s, then later the US Navy.

Edwards:

Oh yeah? Who was that?

Arakawa:

Thomas Rosmond was very interested in medium range prediction, and they were interested in our model, so they supported us. That was a very complicated administrative process. Rather than they give us money directly, they gave money to NSF so it may supplement the NSF grant to us. That continued for a few years. But the changing money from one agency to another took a long time and was very complicated. Now this is the Department of Energy that was one of the big supporters, so that reflects political situations. [Chuckles] But for us, what we were doing, it has all been consistent.

Edwards:

What about Yale Mintz, did he have involvements in political issues that you’re aware of?

Arakawa:

He was a very active person in many, many ways. I don't know that he can be called political. He is a very aggressive person. From the middle of 1970s he regularly took sabbatical and went to Israel at least for one quarter, and then he was on leave of absence, and he finally resigned, when, in the early 1980s and moved to NASA. But NASA, there’s Milton Halem. I think Mintz was relatively free from the politics in NASA, if there is any. So I think he enjoyed just doing research there. So I cannot answer your question. He is a very outgoing, aggressive person, but I hesitate to say he is political.

Edwards:

Did he serve on a lot of panels, like the one you mentioned of the NAS. That seems like something he would likely be invited to do. Do you remember any of those that may have been important?

Arakawa:

He had contact with a very broad variety of people, but I don’t characterize his activity as political.

Edwards:

A couple more questions about issues that might were politicized, but I’m just interested in your general involvement with these, if any. When was the nuclear winter issue? Because your colleague here, Richard Turco, was involved with that with the Turco-Sagan group that came up with that idea in the early 1980s. Did you participate in that debate at all? Did you have discussion with him? Because one of the criticisms that is often leveled against the Turco-Sagan group is they used a very simple climate model to get their projection of a nuclear winter, and when the same kinds of forcings were applied to GCMs it didn’t show nearly the same level of effect.

Arakawa:

That kind of work can be criticized anyway. I may be the one who knows the GCM very well, but that includes its weakness. I don’t object to that research; in fact I’m interested in seeing that. But that’s something I am not interested in myself because I am aware of so many problems. That represents our current state of art, so that’s fine. Such things can be said to nuclear winter, maybe to a different degree from the CO₂problem. And of course real climate change cannot be said without involving the ocean, and involving the ocean raises some bad problems. But I think what really matters is the tendency for that kind of thing to happen, but I don't know how correct those numbers are. Yeah, it’s very easy to criticize. In that sense, even the most advanced GCM right now can be criticized. It’s the state of art. And most GCMs have the difficulty when coupled with the ocean if the flux is not adjusted so that it shows a very serious weakness. Flux adjustment may be a good way to stay away from that weakness for the time being, but of course that’s not really the objective of the model to stay away from the problem, not solving the problem. So although Turco is a colleague here and he is a very good friend, I really don’t know the details of that work. Of course I know about it, but he did that work before he came here. Actually impact will be very different if we consider the ocean. Involving the ocean can delay the time depending on how deep the ocean that is involved is, it can delay. Warming may take place eventually, just answered in this way. Now in the nuclear winter case it is a bit more subtle. It is a very much shorter time scale. Then if the ocean delays significantly, then the overall effect may change significantly, quantitatively speaking.

Edwards:

Another issue is the ozone depletion issue, and that you said yesterday became part of your model, you put ozone chemistry into the model in the CIAP study.

Arakawa:

But it is not doing a very good job. That is very simple, and we put in that to satisfy DOT’s interest in the early 1970s, we put in that. We have been using that, but that is so poor and too simplified that it’s not doing good for the model; it is rather doing harm. Because of the oversimplification, the ozone prediction is not very good, and that hurts the model. So if we just prescribe ozone, giving up prediction of ozone and prescribe the ozone, then the model results are better. But of course versions that we are using are not prescribing the ozone at least right now. That is covered by Turco and Mechoso, I think that Turco and his people are doing a lot. There are many traces of Turco’s stuff in the GCM.

Edwards:

The models you’ve done haven’t prescribed ozone.

Arakawa:

No, his people are using our GCM as well as other models. A man working for him is working with the GCM and putting traces on the chemistry, and also the ocean. That is a big project of Mechoso, Turco and myself. I think that is sponsored by NASA. It is a very big project. If you are interested in very recent kind of status, it may be well for you to see Mechoso because I’m trying to phase out from that.

Edwards:

The problem is I’m interested in so much! For this project I want to focus on the historical things. But Mechoso sounds like he has been important for a long time in other areas, too.

Arakawa:

Yes, he is right.

Edwards:

I have one last question, and this is sort of grab-bag question, and it’s because this invention will become part of the historical record of your work. So I wanted to ask you what, in your own view, are the most important things that you have contributed to climate science? Obviously there is an awful lot, but if you had to pick three or four things that you think are your very most important contributions, what would you say?

Arakawa:

It’s hard to say. In terms of scientific literature is concerned, I mentioned the 1966 paper, so-called the Jacobian paper. It deals with computational aspects but nonlinear, but not really a mathematical approach but from the point of view of the physics of discrete system. That has become a famous paper, and it is very broadly cited, not only meteorology — in journals I never heard of cited that paper in engineering and wherever fluid is involved. That is a very broad citation. And number of citations still remains constant in time. Number of citations is constant also for the Arakawa-Schubert paper of 1974. As I told you, that is not a very difficult paper to understand, at least I didn’t think so. Even before that was published I gave an oral presentation that became already kind of famous. Well, it is famous, but that does not mean all people agreed. There are two types of reactions to that paper: one is from the modelers and theoreticians that say it’s too complicated. There are criticisms from the observationalists that say it is too simple — clouds are not that simple. Then of course this paper might have tried to make some simple things unnecessarily complicated. That’s a very negative way of looking at it. Of course I don’t look at it that way, but I’m sure some people looked at it that way. Relatively recently that’s changing partly because more people recognize the problem is complicated and began to appreciate the depth of this paper. Before, following this approach was out of question for operational models — it’s too expensive. But the computer became faster, and also some people worked on this to make the model more efficient without losing essence. There are many names: Relaxed Arakawa-Schubert, Simplified Arakawa-Schubert, and Prognostic Arakawa Schubert… Some of those are used in the operational models. The number of operational models using it is increasing recently, so it’s one reason the citation of this paper is even increasing now, so maybe I published that paper too early. Another paper, you have Arakawa-Lamb, which is basically a description of the UCLA GCM published in 1977. But inside there were many things new at that time. Some people call that the bible of modeling. So that paper is still being cited. Especially the A, B, C, D, E grid. The choice of that grid influences adjustment mechanisms, and also influences the existence of the computational model. There are other papers, but those three are the most frequently cited. But I guess my overall contribution to the society is not just through individual papers but collectively. The model here had a broad influence at institutions, and I think it is my main contribution.

Edwards:

Well, great. Thank you so much. This has been incredibly interesting.