Home

Description

Machine learned force fields (MLFFs) are extremely exciting tools capable of simulating materials systems from aa single atom to billions of atoms with quantum mechanical accuracy over nanoseconds to milliseconds. In principle, entire processes, for example catalytic reactions, can be modeled in one simulation. However, a method for the validation of the bonding dynamics that MLFFs produce is lacking. This project aims to address this challenge by developing a pipeline that links MLFF-driven dynamic simulations with simulated and experimental spectroscopy.

X-ray absorption spectroscopy (XAS) is appealing for this task because the XAS spectrum encodes "snapshots" of all atomic configurations around the absorbing atoms in the material. By calculating the XAS spectrum on molecular dynamics trajectories, it is possible to benchmark the MLFF that drives the MD trajectories. If the simulated XAS matches experiment, it is evidence that MLFF is accurately capturing the bonding dynamics of the material.

Methods

MLFF validation workflow
Figure 1. The pipeline. A machine-learned force field (MLFF) is trained on a high performance computing cluster using the FLARE framework. The MLFF is then used to drive a molecular dynamics (MD) simulation for the desired system. The resulting MD trajectory is transferred to a local or cloud-based computer for analysis via the data analysis package. This package is used to extract mechanistic information that can be correlated to the experimental analysis such as the evolution of coordination numbers or the radial distribution function (RDF) over time, and parameters that can be extracted the from the RDF such as mean and variance per coordination shell. The MD trajectory is also processed and sent to a HPC for spectral simulation (in this case, the Institutional Cluster of the Center for Functional Nanomaterials at Brookhaven National Lab). The simulation workflow is highly optimized for speed and storage efficiency in order to deal with large MD trajectories. The simulated spectra are then compared to experimental data to validate the MLFF.

Why does this work?

Key Takeaways

Now that a method exists for validating MLFF dynamics, we are able to screen MLFFs for their ability to capture atomic mechanisms.