Publication in Chemical Science (2026)

11 02 2026

A research article has been published in Chemical Science (Royal Society of Chemistry), with PhD student Mikołaj Martyka as the first author and Dr Joanna Jankowska as the corresponding author. The article entitled “Gradients not needed: ML-driven propagation of nonadiabatic molecular dynamics without reference gradients” was carried out within the PRELUDIUM grant (National Science Centre, Poland) in collaboration with Texas Tech University (USA) and Xiamen University (China).

Nonadiabatic molecular dynamics (NAMD) simulations are a key tool for studying photochemical processes such as photoswitching and photoisomerization. Traditionally, these methods require access to analytical expressions for energy gradients, which significantly limits the choice of quantum chemical methods that can be applied in dynamical simulations.

In this work, the authors presented a novel machine learning approach that enables NAMD simulations without access to reference energy gradients. The method is based on fine-tuning the foundational OMNI-P2x model on excited-state energies alone, followed by the use of automatic differentiation to obtain atomic forces. This approach was validated on model photochemical systems, demonstrating good agreement with reference simulations across multiple levels of theory.

A key achievement of this work is enabling, for the first time, dynamics simulations at the QD-NEVPT2 level – an advanced multireference method for which analytical gradients are not available. The method was applied to study the photoisomerization of trans-azobenzene, a prototypical molecular photoswitch, yielding the most accurate simulations of this photochemical process to date.

This research was carried out within the National Science Centre project (PRELUDIUM 2025/57/N/ST4/03587), led by Mikołaj Martyka.

Bibliographic data and link to the article:

M. Martyka, J. Jankowska, H. Lischka, P.O. Dral, Chemical Science 2026, Advance Article

https://doi.org/10.1039/D5SC09557C