Emulating Nuclear Spatio-Temporal Fields Across Model Parameters Using Operator Learning

Advances in nuclear science are driven by the theory-experiment cycle: the process of theorists developing models to explain experimental results, which in turn informs and motivates new experiments. This cycle requires new experimental data and theoretical model evaluations that are costly to obtain. Tasks like model calibration or uncertainty quantification can become computationally prohibitive when repeated evaluation of complex models is required. This bottleneck has motivated the recent development of surrogate models (or emulators) capable of trading a minimal drop in calculation accuracy for massive reductions in computational cost. With next-generation experimental facilities coming online, such as the Facility for Rare Isotope Beams (FRIB), there is a particularly high demand for such emulators capable of quickly and robustly analyzing new data to capitalize on the discovery potential afforded by the new machines.

In this work, we adapt a neural operator paradigm developed within the fluid dynamics community to efficiently emulate Time Dependent Density Functional Theory simulations of nuclear densities/currents of a vibrating calcium nucleus. We present preliminary results that our model captures the time-dynamics across 2 varying parameters of the Skyrme-type energy-density functional, providing evidence that ML can help perform previously impossible uncertainty quantification studies on theoretical models of the real-time dynamics of atomic nuclei.