ACS Fall 2023 Keynote Talk - Machine Learning; Learning Humans

Prof Ong gave a keynote talk at the ACS Fall 2023 meeting on Aug 14 2023. In this talk, he shared three key lessons learned from the past 5 years of working on machine learning in materials science. They are:

Lesson 1: Representation is key. Only representations that encode structure can achieve the necessary accuracy for most properties. Graphs are one such type of representation. The unique feature of MEGNet is the global state variable, which provides additional flexibility to address data limitations.

Lesson 2: Models do learn chemistry. Latent features, such as elemental embeddings, encode chemical information. Such chemical information can be reused for more data-efficient models, e.g., as encodings or part of a multi-fidelity model.

Lesson 3: Keeping as much science as possible. Direct structure-to-property models should be a last resort. Where possible, use machine learning on the smallest "step" in the materials property computation workflow (e.g., modeling the potential energy surface) and rely on well-established phenomenological laws to derive properties.