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

Area of Science:

• Computational Chemistry
• Materials Science
• Machine Learning

Background:

• Nonadiabatic molecular dynamics (NA MD) is crucial for studying far-from-equilibrium processes.
• Calculating excitation energies and NA couplings in NA MD is computationally intensive.
• Existing machine learning models struggle with the complex NA Hamiltonian's dependence on atomic geometry.

Purpose of the Study:

• To develop a computationally efficient method for NA MD simulations.
• To accurately interpolate the NA Hamiltonian using machine learning.
• To extend the accessible timescales for NA MD simulations in materials.

Main Methods:

• Employed bidirectional long short-term memory networks (Bi-LSTM) to interpolate the NA Hamiltonian in the time domain.
• Applied the multiscale approach to three metal-halide perovskite systems.
• Integrated Bi-LSTM with ML force fields for extended simulations.

Main Results:

• Achieved a two-order-of-magnitude reduction in computational cost compared to direct ab initio calculations.
• Obtained reasonable charge trapping and recombination times with NA Hamiltonian sampling every half picosecond.
• Extended NAMD simulation times from picoseconds to nanoseconds, capturing slow dynamics.

Conclusions:

• The Bi-LSTM-NAMD method offers significant computational savings and outperforms previous models.
• This approach accurately captures both fast and slow timescales in NA MD.
• The methodology enables nanosecond-timescale simulations, vital for understanding complex material behaviors, including defects and interfaces.