AI for Materials on Demand | AI for Science

At the AI for Science event, Aron Walsh (Professor of Materials Design at Imperial College London and CSO at CUSP ai) discusses how machine learning is accelerating materials discovery by bridging the predictive power of quantum mechanics with the speed of classical simulations. He outlines key components of materials AI datasets, automation, models, and optimisation—while emphasising constraints like element availability, toxicity, and sustainable computing. Walsh describes approaches to representing chemical systems (hand-built features, learned embeddings, and foundation models), methods to navigate vast chemical spaces, and the shift from screening pipelines to inverse design using generative models such as diffusion and latent-space methods with reinforcement learning. He highlights agent-based workflows, CUSP ai’s materials discovery platform, and the DigiBat robotic lab for automated battery data generation, and closes with a UK outlook on materials 4.0, infrastructure, and industry activity. Aron Walsh is Chief Scientific Officer at CuspAI, and Professor of Materials Design at Imperial College London. His research spans computational materials science, with expertise in electronic structure theory and machine learning for materials discovery. He was awarded the EU-40 Prize for his work on the theory of next-generation solar cells, and the RSC Corday-Morgan Prize for contributions to computational chemistry. 00:00 Speaker Introduction 01:03 From Quantum to Classical 03:29 Machine Learning Potentials 05:17 Inverse Design Blueprint 06:56 Constraints and Green Compute 09:11 Encoding Chemistry for ML 13:05 Learned Element Embeddings 17:27 Vast Chemical Search Space 20:27 Rules and Spinel Mapping 22:14 Generative Models for Crystals 27:36 Screening vs Integrated Agents 31:42 Robotic Battery Discovery 33:36 UK Materials AI Outlook 35:26 Closing Thanks