Forecasting nonadiabatic dynamics using hybrid convolutional neural network/long short-term memory network

Abstract

Modeling nonadiabatic dynamics in complex molecular or condensed-phase systems has been challenging especially for the long-time dynamics. In this work, we propose a time series machine learning scheme based on the hybrid convolutional neural network/long short-term memory (CNN-LSTM) framework for predicting the long-time quantum behavior given only the short-time dynamics. This scheme takes advantage of both the powerful local feature extraction ability of CNN and the long-term global sequential pattern recognition ability of LSTM. With feature fusion of individually trained CNN-LSTM models for the quantum population and coherence dynamics, the proposed scheme is shown to have high accuracy and robustness in predicting the linearized semiclassical and symmetrical quasiclassical mapping dynamics of various spin-boson models with learning time up to 0.3 ps. Furthermore, if the hybrid network has learned the dynamics of a system, this knowledge is transferable that could significantly enhance the accuracy in predicting the dynamics of a similar system. The hybrid CNN-LSTM network is thus believed to have high predictive power in forecasting the nonadiabatic dynamics in realistic charge and energy transfer processes in photoinduced energy conversion.

All-Atom Nonadiabatic Semiclassical Mapping Dynamics for Photoinduced Charge Transfer of Organic Photovoltaic Molecules in Explicit Solvents

“…Incredible efforts have been made in recent decades to elucidate the molecular underpinnings of such phenomena using both experimental and theoretical techniques. − For example, two-dimensional electronic spectroscopy provides rich knowledge on ultrafast dynamical information of CT and EET in realistic light-harvesting molecules in complex or condensed-phase systems. − Theoretical description of nonadiabatic dynamics has advanced significantly. Mixed quantum–classical dynamics such as mean-field Ehrenfest , and fewest-switches surface hopping (FSSH) , are the most popular approaches for simulating the nonadiabatic dynamics in molecular or extended nanoscale systems (up to hundreds of atoms) with the help of on-the-fly electronic structure calculations and the recent machine-learning-based acceleration. − …”

Automatic Evolution of Machine-Learning-Based Quantum Dynamics with Uncertainty Analysis

“…With the rapid development of computer artificial intelligence, machine learning (ML) plays more and more important roles in theoretical chemistry, including the construction of the molecular Hamiltonian, − the analysis of trajectory-based molecular dynamics evolution, − the prediction of molecular properties and chemical reactions, − and the design of novel functional materials. , In addition to these, considerable efforts were made to employ the ML approach to simulate dynamics evolutions recently. − …”

“…In these years, several theoretical efforts tried to employ the ML approaches to simulate the quantum dynamics of reduced systems. − Some works tried to build an ML model based on the short-time dynamics and then use it to predict the long-time evolution. These treatments are highly correlated to the transfer tensor method originally proposed by Cao and co-workers .…”

A comparative study of different machine learning methods for dissipative quantum dynamics

“…Recently, CNNs is used to study the excitation energy transfer in Fenna-Matthews-Olson light-harvesting complex. 34,35 Wu et al 122 used a hybrid CNN/LSTM network to predict long-time semiclassical and mixed quantum-classical dynamics of the spin-boson model. Lin et al 123,130 trained a multi-layer LSTM model to simulate the long-time dynamics of spin-boson model and used bootstrap method to estimate the confidence interval.…”

“…Recently, many ML models have been applied to simulate the dynamics of quantum systems. [32][33][34][35][117][118][119][120][121][122][123][124][125][126] We note that ML can also be applied to quantum dynamics in a different context-namely as surrogate models for quantum chemical properties such as potential energies and forces in different electronic states as well as cou-plings between states eliminating the need for expensive (excited-state) electronic structure calculations. [127][128][129] Here we apply ML to propagate a quantum system assuming potential energies are readily available.…”