From short-sighted to far-sighted: A comparative study of recursive machine learning approaches for open quantum systems

Abstract

Accurately modeling open quantum system dynamics is crucial for advancing quantum technologies, yet traditional methods struggle to balance accuracy and efficiency. Machine learning (ML) provides a promising alternative, particularly through recursive models that predict system evolution based on past history. While these models have shown success in predicting single observables, their effectiveness in more complex tasks, such as forecasting the full reduced density matrix (RDM), remains unclear. We extend history-based recursive ML approaches to complex quantum systems, comparing four physics-informed neural network (PINN) architectures: (i) single-RDM-predicting PINN (SR-PINN), (ii) SR-PINN with simulation parameters (PSR-PINN), (iii) multi-RDMs-predicting PINN (MR-PINN), and (iv) MR-PINN with simulation parameters (PMR-PINN). These models are applied to the spin-boson (SB) model and the Fenna-Matthews-Olson (FMO) complex. Our results show that SR-PINN and PSR-PINN, constrained by a narrow history window, fail to capture complex quantum evolution, leading to unstable long-term predictions, especially in nonlinear and highly correlated dynamics. In contrast, MR-PINN and PMR-PINN improve accuracy by extending the forecast horizon, incorporating long-range correlations, and reducing error propagation. Surprisingly, explicitly including simulation parameters such as temperature and reorganization energy in PSR-PINN and PMR-PINN does not consistently enhance accuracy and can even reduce performance, suggesting that these effects are already encoded in the RDM evolution. These findings highlight the limitations of short-sighted recursive forecasting and demonstrate the superior stability and accuracy of far-sighted approaches for long-term predictions.