Guest lectures: Foundational models for materials chemistry

We are excited to invite you to a guest lecture next Wednesday featuring Professors Gábor Csányi and Javier Antorán from the University of Cambridge. They will be presenting their recent work on machine learning for molecular dynamics simulations. Everyone's welcome to attend, and please feel free to extend this invitation to anyone who might be interested.

We look forward to seeing you there.

 

Title: Foundational models for materials chemistry

Date: Wednesday, 2 October 2024

Time: 12:00-12:55

Location: Building 321, room 232 and https://dtudk.zoom.us/j/61321877779?pwd=DfyOZfoCaGn9ajhwoobn0VJorHcoip.1 

A new computational task has been defined and solved over the past 15 years for extended material systems: the analytic fitting of the Born-Oppenheimer potential energy surface as a function of nuclear coordinates. The resulting potentials ("force fields") are reactive, many-body, with evaluation costs that are currently on the order of 0.1-10 ms/atom/cpu core (or about 1-10ms on a powerful GPU), and reach accuracies of a few meV/atom when trained specifically for a given system using iterative or active learning methods. The latest and most successful architectures leverage many-body symmetric descriptions of local geometry and equivariant message passing networks. Perhaps the most surprising recent result is the stability of models trained on very diverse training sets across the whole periodic table. Our recent discovery is that the MACE-MP-0 model that was trained on just ~150,000 real and hypothetical small inorganic crystals (90% of training set < 70 atoms), is capable of stable molecular dynamics at ambient conditions on any system tested so far - this includes crystals, liquids, surfaces, clusters, molecules, and combinations of all of these. The astounding generalisation performance of such foundation models opens the possibility to creating a universally applicable interatomic potential with useful accuracy for materials (especially when fine-tuned with a little bit of domain-specific data), and democratise quantum-accurate large scale molecular simulations by lowering the barrier to entry into the field. Similarly, in the domain of organic chemistry, training just on small molecules and small clusters allows accurate simulation of condensed phase systems, and first principles prediction of quantities such as hydration free energies for the first time.

Gábor Csányi is a professor of molecular modelling at the University of Cambridge and an expert in atomistic simulation, particularly in multi-scale modelling that couples quantum mechanics to larger length scales. He is currently engaged in applying machine learning and other data-intensive techniques to physics, chemistry and material science and, in particular to the problem of deriving force fields (interatomic potentials) from ab initio electronic structure data. Gábor is also interested in statistical problems in molecular dynamics, e.g. in enhanced sampling algorithms that can be used to explore the global configuration space of materials and molecules. He contributes to the running of the Lennard-Jones Center in Cambridge, which brings together modellers who work on the atomic scale.

Javier Antorán is a research fellow in molecular modelling and probabilistic machine learning at the University of Cambridge. His interests span generative models applied to molecular modelling,  probabilistic modelling, approximate inference and information theory. His PhD research focused on scalable probabilistic reasoning with neural network models.