Metallic alloys uniquely combine strength, ductility, and reusability. These properties result from a complex hierarchy of mechanisms that are only partially understood.
Our research pushes the frontier of multi-scale simulation methods for metals, connecting atomic dynamics, microstructure, and mechanical properties.
Current theoretical topics include:
– Misspecification-aware uncertainty quantification for surrogate models
– End-to-end differentiable simulations for error control and inverse design
– Efficient fine-tuning of foundational atomic models
– Data-driven forecasting of long-timescale simulation behaviour
– Massively parallel algorithms for timescale acceleration
Atomic simulations of materials often reduce to performing high-dimensional integration though Markov Chain Monte Carlo (MCMC) with some model of atomic interaction. This is an essential but very expensive step of modern simulation workflows. Recently, we have pioneered "model agnostic" methods, where MCMC is performed without a priori specification of model parameters. This allows us to pre-compute expensive routines and enables simple forward and back parameter propagation. The major benefits are that we can simply propagate uncertainties through MCMC, and include MCMC observables in objective functions for a variety of inverse design tasks.
Another exciting topic we have been looking at is misspecification-aware regression. Misspecification arises when we have finite error against training data, no matter what value of model parameters are chosen. Think of fitting a quadratic polynomial to data generated by a sinusoid. Such situations are ubiquitous in surrogate modelling, and in fact misspecification dominates in the limit of large data (weak epistemic) and low noise (weak aleatoric). One can prove Bayesian inference is blind to misspecification, as the expected loss is only a (Jensen) upper bound to the true generalization error. We have developed a very fast scheme to determine the true, misspecification-aware parameter posterior for deterministic models, which are regularly encountered in scientific computing. There are numerous exciting extensions for error-controlled modelling which we are beginning to explore, falling into the emerging area of "post-Bayes" modelling.
I first read Physics at the University of Oxford followed by a PhD at Imperial College London, in the Physics department but studying stochastic dynamics of materials. After a postdoc in the Theory Division at Los Alamos and an independent research fellowship at the JET Fusion laboratory near Oxford, I was a permanent researcher in the Physics division of the CNRS in Marseille, France, where I retain a visiting position.
My field of "bottom up" materials design has to make predictions in a strongly extrapolative regime, as we are connecting atomic dynamics to engineering properties. Traditionally this required physics-based models, which have sufficient inductive bias to control the extrapolation behaviour, but they typically struggle to reach quantitative accuracy. The diversity of new tools and ideas in the AI/ML space is incredibly exciting for my field as it allows us to blend physics-based models with highly flexible regression models to combine extrapolation robustness with quantitative accuracy, incorporating quantum mechanical simulation data and available experimental observations.