Fast generation of quantum dynamics data using a GPU implementation of the time-dependent Schrodinger equation

Efficient methods for generating samples of wave packet trajectories are needed to build machine learning models for quantum dynamics. However, simulating such data by direct integration of the time-dependent Schrodinger equation can be demanding, especially when multiple spatial dimensions and realistic potentials are involved. In this paper, we present a graphics processor unit (GPU) implementation of the finite-difference time-domain (FDTD) method for simulating the time-dependent Schrodinger equation. The performance of our implementation is characterized in detail by simulating electron diffraction from realistic material surfaces. On our hardware, our GPU implementation achieves a roughly 350 times performance increase compared to a serial CPU implementation. The suitability of our implementation for generating samples of quantum dynamics data is also demonstrated by performing electron diffraction simulations from multiple configurations of an organic thin film. By studying how the structure of the data converges with sample sizes, we acquire insights into the sample sizes required for machine learning purposes.