Machine Learning Accelerated Molecular Dynamics

Active Learning of Bayesian Force Fields

We are developing Bayesian inference and active learning methods to automate the training of many-body machine learned (ML) force fields for complex materials. Using uncertainties derived from Bayesian ML models, we draw structures "on the fly" from molecular dynamics simulations to automatically update the training set of the model. This process does not need human intervention and results in

Example publications:

Neural Equivariant Interatomic Potentials

NequIP showed for the first time that geometric equivariance leads to interatomic potentials with remarkable accuracy and an unprecedented data efficiency [1]. This deep neural network model leverages E(3)-equivariant convolutions over geometric tensors instead of invariant convolutions over scalars. NequIP is able to outperform existing, invariant Machine Learning potentials with up to 1000x fewer reference data.

Allegro is a new equivariant model that combines accuracy with scalability [2], with parallel performance reaching simulation sizes of tens of millions of atoms, at near first-principles accuracy.

We leverage these potentials to study a wide variety of systems at a high level of accuracy, including surface reaction in heterogeneous catalysis, organic molecules, and complex ionic conductors.

[1] Nature Comm 13, 2453 (2022),


Dimension reduction, enhanced sampling

Learning reaction coordinates. We developed a data-driven machine learning algorithm that uses a multitask encoder neural network model to automatically reduce dimensionality of molecular dynamics trajectory data and learn the latent space of collective variables. This approach is promising for the extraction of atomic-level mechanisms governing reactions in complex chemical systems and for estimation of free energy barriers and reaction rates, especially in cases where human intuition is limited.