Neural-network force field backed nested sampling: Study of the silicon
This proof-of-concept project marked the first phase of my PhD with Georg Madsen and demonstrated that nested sampling (NS) can be successfully combined with neural-network force fields (NNFFs), using silicon as a test system. The project came with a number of technical challenges, and working through these helped me a lot to gain the understanding of machine learning potentials and nested sampling that I have today.
One of the main challenges was managing the computational demands of NS, which are substantially higher than in standard molecular simulations. I began with the pymatnest package, developed by my colleague Livia Bartók Partáy, which provides a flexible Python-based framework for atomistic NS simulations. To make it more efficient for my use case, I replaced its original parallelization scheme with a Dask-based implementation, which led to a significant reduction in runtime and ultimately enabled the NNFF-backed simulations.
A second major challenge was obtaining a neural-network force field that provides a globally accurate representation of the potential energy surface (PES), as required by the NS algorithm. After multiple unsuccessful attempts at training my own model, I turned to a carefully curated, published training database for silicon, which provided the needed PES coverage.
With these challenges resolved, we were able to simulate the low-pressure phase diagram of silicon with remarkable accuracy. The simulation reproduced the melting line as well as all experimentally observed solid phases. We also investigated the influence of the exchange-correlation (XC) functional used during NNFF training on the predicted finite-temperature behavior, demonstrating that NS can serve as a useful tool for benchmarking XC functionals with respect to thermodynamic properties.
I might spice up this dry post with some interactive visualizations soon, stay tuned!
Unglert, N.; Carrete, J.; Pártay, L. B.; Madsen, G. K. H. Neural-Network Force Field Backed Nested Sampling: Study of the Silicon