Applications of Euclidean Neural Networks to Understand and Design Atomistic Systems

Explore how Euclidean neural networks revolutionize the understanding and design of atomistic systems in this 56-minute conference talk from Harvard CMSA's Big Data Conference 2025. Learn about the fundamental challenges of applying machine learning to atomic systems like molecules, crystals, and proteins, which are naturally represented by 3D coordinates but are sensitive to rotations, translations, and inversions. Discover how Euclidean symmetry-equivariant Neural Networks (E(3)NNs) specifically address these challenges by faithfully capturing the symmetries of physical systems and operating on scalar, vector, and tensor fields that characterize these systems. Examine the state-of-the-art results E(3)NNs have achieved across various atomistic benchmarks, including small-molecule property prediction, protein-ligand binding, and force prediction for crystals, molecules, and heterogeneous catalysis. Understand how these networks merge neural network design with group representation theory to embed physical symmetries directly into learning processes. Survey recent applications of E(3)NNs to materials design and engage with ongoing debates in the AI for atomistic sciences community regarding the balance between incorporating physical knowledge and maintaining engineering efficiency.