Machine learning accelerated nonadiabatic dynamics simulations of materials with excitonic effects

Area of Science:

• Computational Chemistry
• Materials Science
• Quantum Mechanics

Background:

• Simulating nonadiabatic dynamics in complex materials with excitonic effects is computationally expensive.
• Accurate prediction of excited-state properties and nonadiabatic couplings is crucial for understanding photoinduced processes.

Purpose of the Study:

• To develop an efficient machine learning (ML) methodology for simulating nonadiabatic dynamics.
• To integrate ML models with simplified Tamm-Dancoff approximation (sTDA) for accelerated calculations.
• To enable accurate and fast simulations of photoinduced dynamics in large material systems.

Main Methods:

• Leveraging ML models to predict ground-state wavefunctions from unconverged Kohn-Sham (KS) Hamiltonians.
• Employing ML-predicted KS Hamiltonians for sTDA-based excited-state calculations (sTDA/ML).
• Performing nonadiabatic molecular dynamics simulations using the sTDA/ML approach.

Main Results:

• sTDA/ML calculations show high accuracy for excited-state energies, nonadiabatic couplings, and absorption spectra compared to sTDA/DFT.
• Achieved over 100x speedup in nonadiabatic molecular dynamics simulations for silicon quantum dots and black phosphorus.
• Demonstrated computational savings without compromising the accuracy of simulations.

Conclusions:

• The ML-accelerated sTDA methodology offers a powerful tool for studying complex material dynamics.
• This approach significantly reduces computational cost for large systems, enabling broader research.
• Highlights the potential of ML in advancing computational materials science and photophysics.