Publications

Toward Machine Learning Enhancement of Accelerated Surface Hopping with Scaled Spin–Orbit Couplings

# machine learning # accelerated surface hopping # intersystem crossing # spin-orbit coupling

Surface hopping simulations are well suited for ultrafast processes, such as internal conversion through conical intersections, which typically happen within hundreds of femtoseconds. Intersystem crossing (ISC), the transition between states of different spin multiplicity, is a different story. When the spin–orbit couplings (SOCs) driving it are small, ISC can take hundreds of picoseconds, which is far beyond what we can afford to simulate with high-level electronic structure methods.

If the goal is only to estimate the time constant, there is a trick: the SOCs can be artificially scaled by a constant \(\alpha > 1\) so that ISC happens within the simulated time window. Running several ensembles of trajectories with different scaling factors and extrapolating back to \(\alpha = 1\) then gives the actual time constant. The catch is that every scaling factor requires its own ensemble of trajectories, which quickly becomes expensive.

In this work, we look at how reliable such an accelerated scheme is, using silaethylene (CH2SiH2) as a test case. We trained machine learning (ML) models for the potential energy surfaces, nonadiabatic couplings and SOCs, building on our NAC descriptor and rotate-predict-rotate approaches, and used them to explore where ML can make the whole procedure more robust.

For detailed information, check out the preprint: (Martinka et al., 2026). The ML part of this work was carried out in MLatom.

References

2026

  1. so_scaling.png
    Toward Machine Learning Enhancement of Accelerated Surface Hopping with Scaled Spin–Orbit Couplings
    Jakub Martinka, Mahesh Kumar Sit, Pavlo O. Dral, and Jiří Pittner
    2026 arXiv:2604.25603