Materials science × Artificial intelligence
We develop AI and atomistic simulations to understand how materials evolve, react and function.
Machine-learning potentials (MLPs)
Simulation & discovery
Compare carbon neighbourhoods inside molecules, solutions, polymers and crystals. Then explore how diverse configurations are selected for labelling.
Explore how machine-learning potentials encode atomic environments through symmetry-aware features. Translate, rotate and reflect the neighbourhood.
Change coverage, surface structure and atmosphere. Follow how they reshape a reaction and its activity.
Surface structure
Reaction atmosphere
Choose activity or selectivity. Follow a different hypothesis through simulation, testing and feedback.
Research objective
Follow the workflow
Explore four ways of connecting atomic structure and materials behaviour.
We connect atomistic data, symmetry-aware machine learning and simulation to understand how materials evolve and react. Heterogeneous catalysis is a central focus: a setting where changing structures and surface populations shape reaction pathways, rates and selectivity.
Better data. Better questions.
We generate, sample and select atomic configurations for reactive machine-learning potentials, and study how training data shape accuracy and transferability.
Bond breaking and formation; configuration generation; training-data selection.
Explore this directionSymmetry, built in.
We develop symmetry-aware representations and efficient architectures for atomic interactions, from Cartesian tensors to local-frame models for rotations, reflections and time reversal.
Cartesian tensors; local frames; rotations, reflections and time reversal.
Explore this directionCatalysis in context.
We connect evolving catalyst structures and surface coverage with reaction mechanisms, rates and selectivity through machine-learning potentials, sampling and microkinetic modeling.
Evolving active sites; surface coverage; sampling and microkinetics.
Explore this directionNew possibilities, evaluated.
We explore generative methods and simulation to propose and evaluate promising materials, with the longer-term aim of guiding informative experimental investigations.
Generative candidates; first-principles evaluation; simulation-informed interpretation.
Explore this directionStudies in atomistic data, physical representations and machine learning.
All publicationsProceedings of the 43rd International Conference on Machine Learning
Nature Catalysis
JACS Au
arXiv
Model architectures, pretrained potentials, equivariant operations and training data.
Software & DataA family of equivariant atomistic models built around tensor representations and flexible physical embeddings, including TACE, TECE and magnetic TACE (mTACE).
Pretrained TACE-family models for atomistic simulation and fine-tuning. The model hub and usage guide describe the available releases and their intended applications.
Explore TACE FoundationsA library of equivariant operations for atomistic modeling, bringing together Cartesian tensor tools and local O(2) representation maps, frame transformations and operators.
An interaction-based approach to generating training data for general reactive element-based potentials. The linked studies address transferability and model construction for heterogeneous catalysis.
Explore REICOSeptember 2026
Reactive potentials, data generation and equivariant models.
September 2026
Rotations, reflections and time reversal in atomistic models.
July 2026
Data-centric methods, interatomic potentials and materials simulation.
We welcome enquiries from students and researchers interested in materials science, theoretical chemistry and machine learning.
Research opportunities