Directly learning molecular dynamics!

Molecules are always in motion.This motion can be simulated with molecular dynamics approach but it needs many steps to evaluate and is very slow. Just to demonstrate the pain of propagating MD let’s do it by hand for a couple of iterations: we calculate the nuclear positions at the future time step with some MD propagator, then calculate forces for the new geometry, and repeat this procedure many thousands or millions times.

We can accelerate this procedure with machine learning potentials to calculate forces faster. You can check out how to do it in our tutorials on MLPs and MD.

However, machine learning potentials still do not solve the fundamental problem with dynamics – it is iterative, steps cannot be parallelized, quality depends on time step, trajectories are discrete and not smooth.

That’s why in March 2022, we introduced a novel concept of directly learning dynamics! The idea is very simple but not easy to realize: we predict the nuclear coordinates as a continuous function of time! We can parallelize predictions at different time steps as they do not need to be calculated sequentially. The trajectories also can be obtained with arbitrarily high resolution.

We managed to pull off a difficult task of creating a deep learning model to learn dynamics and predict the trajectories fast. This model is called GICnet and was published in JPCL in 2023. In effect, GICnet models are analytical representation of molecules and chemical reactions in four-dimensional spacetime!

Amazingly, our model can learn dynamics of different molecules and even learns cis-trans isomerization dynamics of azobenzene. The trajectories can be obtained with GICnet orders of magnitude faster than with machine learning potentials. The GICnet model is now incorporated in MLatom which comes with the online tutorial.