Challenges and opportunities in deep learning for sustainable materials design (09/01/25)
Speaker and Affliation:
Dr. Keith Tobias Butler
Associate Professor, Department of Chemistry, University College London
When?
09th January, 2025 (Thursday), 3.00 PM (India Standard Time)
Where
KPA Auditorium, Dept. of Materials Engineering, IISc, Bangalore
Abstract:
The discovery and design of new materials is critical for advancing carbon-emission reducing technologies such as renewable energy and electric vehicles. Experimental discovery of new materials is typically slow and costly, quantum mechanics (QM) calculations have brought computational materials design within reach. However, QM calculations are often limited to relatively small sets of materials, as their computational costs are too great for large-scale screening, this is the case for calculating properties required for new energy materials. New methods in machine learning (ML) and deep learning (DL) have emerged as a powerful complementary tool to QM calculations – learning rules from data calculated from QM and applying cheap, efficient models to explore large chemical spaces. However, several challenges still exist for example, learning from small and limited datasets, obtaining measures of confidence in models and understanding the results of DL models. All these challenges must be addressed to fully realise the power of DL for design of new sustainable materials. In this talk I will give examples of recent work in our group to address these issues, including using unsupervised learning to accelerate the characterisation of battery materials without requiring labelled data[1], building models with reliable uncertainty quantification [2], capable of learning on significantly smaller datasets than regular DL models, learning general models on large and varied datasets [3] and using DL to match experimental and simulated data [4]. Finally, I will also discuss how the latest developments in large language models could help to solve the challenges of crystal structure prediction [5].
-
Versatile domain mapping of scanning electron nanobeam diffraction datasets utilising variational autoencoders npj Computational Materials 9 (1), 14, 2023
-
Entropy-based active learning of graph neural network surrogate models for materials properties The Journal of Chemical Physics 155 (17), 174116, 2021
-
Optimal pre-train/fine-tune strategies for accurate material property predictions npj Computational Materials 10 (1), 300, 2024
-
Using generative adversarial networks to match experimental and simulated inelastic neutron scattering data Digital Discovery 2.578, 2023
-
Crystal Structure Generation with Autoregressive Large Language Modeling Nature Communications 15, 1-16 2024
About the Speaker :
Keith did an undergraduate degree in Chemistry at Trinity College Dublin, graduating in 2004. In 2006 he came to UCL and did a PhD in Computational Chemistry, under the supervision of Dewi Lewis, studying the nucleation and growth of zeolites. Keith then did post doctoral research in the University of Sheffield and the University of Bath, during this time he was mostly working on simulations of interfaces in photovoltaics, with a particular interest in crystalline silicon solar and hybrid halide perovskites. During his time at the University of Bath, Keith became interested in machine learning for the discovery and analysis of new materials. In 2018 he moved to the Rutherford Appleton Laboratory, where he worked at ISIS and helped to start the scientific machine learning group (SciML). In 2022 Keith moved to Queen Mary University of London as a Senior Lecturer in Green Energy Materials. In 2023 he re-joined UCL Chemistry as an Associate Professor. Keith is a deputy editor at npj Computational Materials and sits on the editorial board of Machine Learning Science and Technology. He is also on the ECR board of AIchemy, the UK’s hub for AI in chemistry.