Harnessing the Properties of Equivariant Neural Networks for Materials Science

Explore the cutting-edge applications of Euclidean symmetry-equivariant Neural Networks (E(3)NNs) in materials science and design during this 1-hour 12-minute lecture by Tess Smidt from MIT. Delve into the challenges of representing atomic systems in machine learning and discover how E(3)NNs address these issues by capturing the symmetries of physical systems. Learn about the state-of-the-art performance of E(3)NNs in various atomistic benchmarks, including small molecule properties, protein-ligand binding, and force prediction for heterogeneous catalysis. Gain insights into the combination of neural network operations with group representation theory, understanding why E(3)NNs are more robust, data-efficient, and capable of generalization compared to other neural network types. Examine recent applications of E(3)NNs in materials understanding and design, and explore the expansion of these methods to new domains and data modalities. Conclude by considering open questions regarding the expressivity, data-efficiency, and trainability of methods leveraging invariance and equivariance in the context of AI and scientific research.