Welcome to the Department of Prof. Gábor Csányi

The broad research remit of the Machine Learning in Molecular Dynamics Department is to use machine learning techniques to advance the efficacy of molecular dynamics, both in terms of accuracy and computational efficiency. Richly parametrised functional forms (whether they be large linear models, such as ACE, nonlinear kernel based models such as GAP, or neural networks such as MACE) allow force fields (also known as interatomic potentials) to be derived from electronic structure calculations systematically, approaching the true Born-Oppenheimer potential energy surface in the limit of large training data and large model size. This means that rather than being empirical, “minimal” models, such force fields become true “first principles” methods. In terms of molecular dynamics, they represent many orders of magnitude speed-up over e.g. density functional theory, and in favourable cases, without giving up any significant accuracy for downstream observables. 

The principles of the MLMD technique are well established, and current methodological research is about widening the range of systems which can be treated, e.g. by including explicit electrostatic terms that are charge-self-consistent (representing a coarse-graining of the electron density), or including magnetic degrees of freedom, or targeting molecular interaction terms that are small in magnitude but important for  predicting properties of molecular liquids and solids.

There are particular areas in chemistry that especially benefit from MLMD techniques: complex liquids (water with many ions, organic electrolytes, confined liquids, molten salts), low temperature catalysis, amorphous materials are all examples in which the newly attainable system sizes of a few-nanometers and time scales of many-nanoseconds are both necessary and sufficient to study material processes and gain new insight. 

The Department has subgroups that are also focussed in terms of application area. Alice Allen leads a biomolecular simulation group, and Fabian Berger leads a group that is interested in surfaces and interfaces.