Gradients not needed: ML-driven propagation of nonadiabatic molecular dynamics without reference gradients

Area of Science:

• Quantum chemistry
• Computational chemistry
• Machine learning applications

Background:

• Machine learning (ML) significantly enhances molecular calculation efficiency.
• Nonadiabatic molecular dynamics (NAMD) simulations typically require analytical energy gradients, limiting method choices.
• Gradient computation is a bottleneck for advanced electronic structure methods in dynamics.

Purpose of the Study:

• To develop and validate a gradient-free ML approach for NAMD simulations.
• To enable NAMD simulations using electronic structure methods lacking analytical gradients.
• To advance excited-state dynamics simulations for complex molecular systems.

Main Methods:

• Fine-tuning a foundational ML model (OMNI-P2x) on energies alone.
• Leveraging automatic differentiability to derive forces for ML potentials.
• Validating the gradient-free ML potentials on benchmark systems like fulvene and cyclohexadiene.

Main Results:

• Gradient-free ML potentials accurately reproduced NAMD populations and dynamics for AIQM1/MRCI, CASSCF, and MRSF-TDDFT.
• Enabled NAMD simulations at the QD-NEVPT2 level for the first time.
• High accuracy was achieved in simulating cyclohexadiene photoinduced ring-opening dynamics.
• Fully dimensional excited-state simulations of *trans*-azobenzene photoisomerization were performed at unprecedented levels of theory.

Conclusions:

• Gradient-free ML potentials successfully overcome the limitations of analytical gradient requirements in NAMD.
• This method significantly expands the scope of accessible high-level electronic structure methods for excited-state dynamics.
• The approach sets a new standard for simulating complex photochemical processes.