Reproduce “Excited-state nonadiabatic dynamics in explicit solvent using machine learned interatomic potentials” on OpenResearch

About this paper

Maximilian X. Tiefenbacher, Brigitta Bachmair, Cheng Giuseppe Chen, Julia Westermayr, Philipp Marquetand, Johannes C. B. Dietschreit, Leticia González

This work introduces a machine learning/molecular mechanics (ML/MM) approach for simulating excited-state nonadiabatic dynamics in explicit solvent, significantly reducing computational cost compared to traditional quantum mechanics/molecular mechanics methods. The framework, tested on furan in water, successfully reproduced electronic population dynamics and identified an energy-gap-weighted error metric as a predictor of simulation accuracy.

Key insights

• The augmented loss function, which incorporates electric field-dependent gradients, is essential for numerical stability and accurate force reproduction in ML/MM electrostatic embedding.
• Energy-gap-weighted Root Mean Square Error (wRMSE) for forces is a robust predictor of an ML model's performance in reproducing nonadiabatic dynamics, particularly near conical intersections.
• Data collection strategies for ML potentials in dynamics should prioritize active learning to sample critical regions of the potential energy surfaces efficiently, rather than relying on standard QM/MM trajectories that oversample correlated frames.

Summary by alphaXiv.