Predicting the future of excitation energy transfer in light-harvesting complex with artificial intelligence-based quantum dynamics

Abstract

Exploring excitation energy transfer (EET) in light-harvesting complexes (LHCs) is essential for understanding the natural processes and design of highly-efficient photovoltaic devices. LHCs are open systems, where quantum effects may play a crucial role for almost perfect utilization of solar energy. Simulation of energy transfer with inclusion of quantum effects can be done within the framework of dissipative quantum dynamics (QD), which are computationally expensive. Thus, artificial intelligence (AI) offers itself as a tool for reducing the computational cost. We suggest AI-QD approach using AI to directly predict QD as a function of time and other parameters such as temperature, reorganization energy, etc., completely circumventing the need of recursive step-wise dynamics propagation in contrast to the traditional QD and alternative, recursive AI-based QD approaches. Our trajectory-learning AI-QD approach is able to predict the correct asymptotic behavior of QD at infinite time. We demonstrate AI-QD on seven-sites Fenna–Matthews–Olson (FMO) complex.

Four-dimensional spacetime atomistic artificial intelligence models

“…Similarly, we have recently demonstrated that for some known asymptotic limits, ML can be trained to predict quantum dynamics up to the infinite future time, but this was done only for a model Hamiltonian and not for nuclear positions. 28 Predictions made with the 4D-A 2 I model for H2 are practically instantaneous while propagating sequential BOMD of even such a small system as H2 up to 15 fs with 3D-MLP and a small time-step of 0.005 fs took longer than 30 min; propagating with full configuration interaction QM method would take several months on a desktop computer (see Methods). The ultimate goal of a complete 4D-A 2 I model is to describe an atomistic system across infinite 4D spacetime, i.e., predict dynamics trajectories for any combination of initial conditions and any time.…”

One-shot trajectory learning of open quantum systems dynamics

“…Recently, ML has been successfully applied to accelerate dynamics propagation of open quantum systems. [10][11][12][13][14][15][16][17][18][19] ML-based approaches adopted in the literature so far, can be divided into two categories: recursive [13][14][15][16] and non-recursive [17,18] approaches. ML-based recursive approaches are quite successful but there are some downsides of them; first, iterative propagation is inherently slow and may lead to error accumulation and second, they need a short time-trajectory generated with traditional quantum dynamics methods.…”

“…[17] Recently, we have proposed an alternative non-recursive, artificial intelligence-based quantum dynamics (AI-QD), approach learning dynamics trajectories as a function of time and simulation parameters, the reorganization energy λ, characteristic frequency γ, and temperature T . [18] This makes all time-steps independent from each other which allows us to predict system's state at any arbitrary time, without the need of propagating the trajectory.…”

“…In addition, the time and memory requirements for training are rather high because of the way the data for supervised learning is prepared: AI-QD is trained on as many points as there are time steps in all training trajectories, i.e., training set contains millions of points. [18] Here we propose a one-short trajectory learning (OSTL) quantum dynamics approach, which is a much more efficient strategy for both learning and trajectory propagation than AI-QD. Similar to AI-QD, OSTL takes the simulation parameters as input but then in contrast to AI-QD, OSTL directly evaluates the entire trajectory as a multi-output prediction by a convolutional neural network (CNN) (Fig.…”

“…Our new strategy only requires up to 50 milliseconds for 2.5 ps long trajectory for the excitation energy transfer in the seven-sites FMO complex [2,3] which was also used in previous ML studies. [17,18] OSTL also has significantly reduced computational requirements for training as a multi-output CNN is trained on the number of points equal to the number of training trajectories.…”