Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Bidirectional Long Short-Term Memory Networks

Essential for understandingfar-from-equilibrium processes,nonadiabatic(NA) molecular dynamics (MD) requires expensive calculations of theexcitation energies and NA couplings. Machine learning (ML) can simplify computation; however, the NA Hamiltonian requires complex ML modelsdue to its intricate relationship to atomic geometry. Working directlyin the time domain, we employ bidirectional long short-term memorynetworks (Bi-LSTM) to interpolate the Hamiltonian. Applying this multiscaleapproach to three metal-halide perovskite systems, we achieve twoorders of magnitude computational savings compared to direct ab initiocalculation. Reasonable charge trapping and recombination times areobtained with NA Hamiltonian sampling every half a picosecond. TheBi-LSTM-NAMD method outperforms earlier models and captures both slowand fast time scales. In combination with ML force fields, the methodologyextends NAMD simulation times from picoseconds to nanoseconds, comparableto charge carrier lifetimes in many materials. Nanosecond samplingis particularly important in systems containing defects, boundaries,interfaces, etc. that can undergo slow rearrangements.