Transfer learning and multi-fidelity modeling of laser-driven particle acceleration

Abstract

Computer models of intense, laser-driven ion acceleration require expensive particle-in-cell simulations that may struggle to capture all the multi-scale, multi-dimensional physics involved at reasonable costs. Explored is an approach to ameliorate this deficiency using a multi-fidelity framework that can incorporate physical trends and phenomena at different levels. As the basis for this study, an ensemble of approximately 8000 1D simulations was generated to buttress separate ensembles of hundreds of higher fidelity 1D and 2D simulations. Using transfer learning with deep neural networks, one can reproduce the results of more complex physics at a much lower cost. The networks trained in this fashion can, in turn, act as surrogate models for the simulations themselves, allowing for quick and efficient exploration of the parameter space of interest. Standard figures-of-merit were used as benchmarks such as the hot electron temperature, peak ion energy, conversion efficiency, and so on. We can rapidly identify and explore under what conditions differing fidelities become an important effect and search for outliers in feature space.