Uncovering the hidden labor behind maintaining a dataset

Everyone wants a robust dataset to train their AI and ML models, but the creation and upkeep can require time and resources that often go unseen by users.

Richardella et al. developed the Lifetime Sample Tracking (LiST) platform and documented how his team categorizes and shares material data from the 2D Crystal Consortium Materials Innovation Platform (2DCC-MIP) at Penn State University.

“We realized pretty early on that the standard way that labs manage their data is kind of haphazard, and [they] tend to lose a lot of stuff,” said author Anthony Richardella. “Getting people working on a common set of standards is probably the biggest inertia — once you push past that, that’s when the gains start to be visible.”

The LiST platform is comprised of over 20,000 samples, and it gives researchers access to characteristics of 2D materials such as growth parameters and material properties.

Richardella and his team began by creating common tracking IDs for all the samples, which previously were handled differently by different labs. That first step allowed them to organize what data went with which sample.

The dataset also includes both “good” and “bad” samples, something that is valuable for training models. Richardella said this was a happy accident — users already have to log their samples when they begin work, so it takes no extra time to document a process that doesn’t work.

The data from the LiST platform is already being used to train existing models, like imaging algorithms that can be trained to recognize topographic features of 2D materials.

Next, Richardella and other collaborators will expand the capabilities of 2DCC-MIP and LiST as a part of the National Science Foundation’s initiative to create a network of programmable cloud laboratories.

Source: “Lifetime sample tracking (LiST): A data platform for materials science,” by Anthony Richardella, Isaiah A. Moses, Konrad Hilse, Frank Santaguida, Kevin Dressler, Ric Wilburn, Saiyyam Kochar, Wesley F. Reinhart, Adri C. T. van Duin, Nitin Samarth, Vincent H. Crespi, and Joan M. Redwing, APL Machine Learning (2026). The article can be accessed at https://doi.org/10.1063/5.0344984 .