Increasing Efficiency of Nonadiabatic Molecular Dynamics by Hamiltonian Interpolation with Kernel Ridge Regression

Abstract

Nonadiabatic (NA) molecular dynamics (MD) goes beyond the adiabatic Born−Oppenheimer approximation to account for transitions between electronic states. Such processes are common in molecules and materials used in solar energy, optoelectronics, sensing, and many other fields. NA-MD simulations are much more expensive compared to adiabatic MD due to the need to compute excited state properties and NA couplings (NACs). Similarly, application of machine learning (ML) to NA-MD is more challenging compared with adiabatic MD. We develop an NA-MD simulation strategy in which an adiabatic MD trajectory, which can be generated with a ML force field, is used to sample excitation energies and NACs for a small fraction of geometries, while the properties for the remaining geometries are interpolated with kernel ridge regression (KRR). This ML strategy allows for one to perform NA-MD under the classical path approximation, increasing the computational efficiency by over an order of magnitude. Compared to neural networks, KRR requires little parameter tuning, saving efforts on model building. The developed strategy is demonstrated with two metal halide perovskites that exhibit complicated MD and are actively studied for various applications.

Interpolating Nonadiabatic Molecular Dynamics Hamiltonian with Inverse Fast Fourier Transform

“…The sampled data capture this fluctuation, while the iFFT interpolation slightly overestimates the fluctuation magnitude. In comparison, a NN with a comparable amount of training data underestimated somewhat the fluctuation magnitude, while a KRR model reproduced this peak accurately using twice more training data . One expects that low sampling should underestimate peak amplitudes, which is the case for other NAC peaks in Figure .…”

“…In comparison, a NN with a comparable amount of training data underestimated somewhat the fluctuation magnitude, 30 KRR model reproduced this peak accurately using twice more training data. 32 One expects that low sampling should underestimate peak amplitudes, which is the case for other NAC peaks in Figure 3. The unusual behavior of the iFFT prediction of the large fluctuation in the CBM−trap NAC at the early time requires further investigation.…”

“…In this letter, we demonstrate that an efficient interpolation of the NAMD Hamiltonian generated under the CPA can be achieved by inverse fast Fourier transform (iFFT). This strategy is notably simpler than the use of NNs 30 and KRR, 32 while at the same time the needed training data set is even smaller. The iFFT method requires much less time and effort than training a ML model, while it gives reliable NAMD results that converge to the ab initio calculations.…”

Fluctuations at Metal Halide Perovskite Grain Boundaries Create Transient Trap States: Machine Learning Assisted Ab Initio Analysis

“…Most of the previous theoretical research on GBs has been based on short, few picosecond (ps) molecular dynamics (MD) trajectories or even just optimized structures due to computational limitations imposed by ab initio methodologies. ,− ,, Recently, machine learning (ML) methods have gained immense popularity as promising tools to overcome the computational cost of ab initio methods to predict material properties , and structures, construct ML force fields (FF), and screen promising candidates for a variety of applications . With the help of ML FFs generation of long MD trajectories with near ab initio accuracy is now feasible, paving the way for studies of phenomena that occur over long time scales, such as anharmonic lattice dynamics, phase transitions, and chemical dynamics. − A ML FF modeling of perovskite GBs can capture the diversity of atomic configurations and their impact on the electronic structure.…”

Extending the Time Scales of Nonadiabatic Molecular Dynamics via Machine Learning in the Time Domain

“…Mangan et al 47 explored the unsupervised ML approaches to establish correlations between the structural features of condensedmatter systems such as lead halide perovskites and their NACs and band gaps, although no use of ML in the NA-MD simulations has been reported. More recently, Wu et al 48 utilized the KRR approach to conduct the NA-MD calculations of perovskite systems using several internal degrees of freedom as the inputs to the ML model to predict the NACs and energy gaps.…”