Machine learning inspires contemporary chemicals research

Through image recognition and reinforcement learning techniques, machine learning is utilized in a wide variety of diverse fields. The existing methods for structure search rely heavily on random elements. As problems tend to become more complex, it is less likely they can be solved by chance as done with Monte Carlo methods and evolutionary algorithms.

Atomistic structure learning applies the concept of image recognition into the field of quantum chemical guided materials search. The identification of atomic coordinates in a structure now may be determined by a self-trained machine learning agent in a few thousand attempts and done in a couple of days.

After first explaining the stages of an atomistic structure learning algorithm (ASLA) -- building, evaluating and learning -- Hammer et al. demonstrate successful atomistic reinforcement learning through the construction of pristine graphene, organic compound formation and surface oxide structure by an ASLA agent.

“We combine reinforcement learning with computational chemical physics and demonstrate how the computer may teach itself basic rules governing structure formation of small, planar organic molecules and two-dimensional crystals and surfaces,” Hammer said.

An ASLA agent learned how to build a sheet of pristine graphene in stages by first building a messy structure and then learning from its errors. After 1,000 attempts, the agent realized bond directions must alternate and then constructed the honeycomb pattern of graphene. Through transfer learning, where knowledge is acquired for one problem and applied to another, the process can be sped up, thereby showing signs of having acquired artificial intelligence.

ASLA will save researchers time, taking tasks representing time-consuming intellectual challenges (such as guessing the structure of matter) and complete them in a matter of days instead of years.

Source: “Atomistic structure learning,” by Mathias S. Jørgensen, Henrik L. Mortensen, Søren A. Meldgaard, Esben L. Kolsbjerg, Thomas L. Jacobsen, Knud H. Sørensen, and Bjørk Hammer, The Journal of Chemical Physics (2019). The article can be accessed at https://doi.org/10.1063/1.5108871 .