Highly efficient nuclear population transfer through physics-informed neural networks

Nuclear coherent population transfer (NCPT) offers numerous potential applications, particularly in next-generation nuclear clocks and nuclear batteries. However, the realization of high fidelity, fast operation, and low energy consumption in NCPT remains so far challenging. Here, we employ physics-informed neural networks (PINNs) to the population transfer in an open three-level nuclear system with spontaneous emission. The method embeds the system's control equations and boundary conditions into the loss function, thereby enabling the automatic learning of optimal laser pulse sequences that drive highly efficient population transfer. We take a short-lived excited state of $^{172}\mathrm{Yb}$ and a long-lived state of $^{229}\mathrm{Th}$ as representative examples, and systematically compare the performance of the PINNs approach with three conventional control strategies. We show that PINNs can achieve higher transfer efficiency with smaller pulse areas and shorter durations across different lifetime regimes. Our results provide a new perspective to overcome the lifetime limitation and enhance the efficiency of nuclear state transfer.