Computational experiments with machine learning to simulate photoinduced processes

Computations based on atomistic-level electron-phonon coupled investigation of dynamical behavior require the use of time dependent density functional theory (DFT) combined with non-adiabatic molecular dynamics (NAMD) simulations [1,2]. These simulations are intensive and expensive for larger systems and/or when the dynamics exceeds few to several tens of picoseconds (ps). Coupling with advanced machine learning (ML) techniques can reduce the time/cost by orders of magnitude, while retaining the ab-initio accuracy. As a testing system, we considered 2D tin-halide perovskites (PEA)₂MA_n-1Sn_nI_3n+1 (with n=1,4) single crystals with tunable quantum well thickness (n) [3]. We adopted the ML mapping approach developed by Akimov's group [4]. To benchmark, we first performed complete calculations on the smallest system using DFT/B3LYP. For the recombination process, 20% training dataset is found to yield ML results in good agreement with DFT, while speeding up the NAMD calculations by a factor of 5. After confirming the success with 20% training, we used the ML model to predict energies and non-adiabatic couplings for the three remaining systems to compute the population decay and recombination times. The length of the organic chain is found to strongly affect the dynamics. The methodology shows promise to apply for NAMD of, namely, Squaraine molecules and Fullerene derivatives.

References:

[1] “Ultrafast transfer and transient entrapment of photoexcited Mg electron in Mg@C₆₀”; Madjet et al., Phys. Rev. Lett. 126, 183002 (2021).

[2] “Ultrafast nonadiabatic electron dynamics in photoexcited C₆₀: a comparative study among DFT exchange-correlation functionals”; Ali et al., J. Phys. Chem. A 129, 2123 (2025).

[3] “Photo-excited carrier behaviors of two-dimensional tin halide perovskite single crystals”; Li et al., Cell Rep. Phys. Sc. 5, 102020, (2024).

[4] “Machine-Learned Kohn–Sham Hamiltonian Mapping for Nonadiabatic Molecular Dynamics”; M. Shakiba and A.V. Akimov, J. Chem. Theory Comput. 20, 2992 (2024).