Nature Communications (Nov 2021)

Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting

  • Stephan Thaler,
  • Julija Zavadlav

DOI
https://doi.org/10.1038/s41467-021-27241-4
Journal volume & issue
Vol. 12, no. 1
pp. 1 – 10

Abstract

Read online

In machine learning approaches relevant for chemical physics and material science, neural network potentials can be trained on the experimental data. The authors propose a training method applying trajectory reweighting instead of direct backpropagation for improved robustness and reduced computational cost.