Related Experiment Video
Updated: Jul 16, 2026
10:52
Direct Imaging of Laser-driven Ultrafast Molecular Rotation
Published on: February 4, 2017
XMCQDPT2-fidelity transfer-learning potentials and a wavepacket oscillation model for ultrafast photodynamics
Ivan V Dudakov1,2, Pavel M Radzikovitsky1, Dmitry S Popov1
1Department of Chemistry, Lomonosov Moscow State University, Leninskie Gory 1/3, Moscow 119991, Russia.
The Journal of Chemical Physics
|July 14, 2026
Summary
Transfer learning effectively trains machine-learning interatomic potentials for simulating photochemical reactions. This approach significantly reduces the need for expensive quantum chemistry data, enabling accurate modeling of nonadiabatic dynamics.
Area of Science:
- Computational Chemistry
- Quantum Mechanics
- Photochemistry
Background:
- Accurate simulation of photochemical reactions requires modeling nonadiabatic transitions.
- Machine-learning interatomic potentials (MLIPs) offer efficiency but face challenges in training for excited states due to data costs.
Purpose of the Study:
- To explore and compare strategies for developing MLIPs that achieve high accuracy for nonadiabatic molecular dynamics.
- To identify the most efficient training methodology for MLIPs targeting excited-state simulations.
Main Methods:
- Systematic comparison of single-state, multi-output, multi-state, transfer learning (TL), and delta-learning architectures for MLIPs.
- Validation of the optimal methodology on the methaniminium cation using XMCQDPT2/SA(3)-CASSCF(12,12) level calculations.
- Development of a wavepacket oscillation model to analyze excited-state population dynamics.
Main Results:
- Transfer learning from CASSCF to XMCQDPT2 demonstrated the best balance of accuracy and computational efficiency.
- The TL approach significantly reduced the required expensive reference data.
- The methaniminium cation's complete photodissociation landscape was accurately captured, revealing S1 branching and H2-loss pathways.
- TL models produced distinct ultrafast population dynamics compared to randomly initialized models.
Conclusions:
- Transfer learning is a highly effective strategy for developing accurate and efficient MLIPs for nonadiabatic molecular dynamics.
- The developed methodology enables comprehensive simulation of photochemical reaction pathways and dynamics.
- The wavepacket oscillation model provides quantitative insights into state- and channel-specific lifetimes, linking quantum probabilities to classical rate constants.
