Predictive Scale-bridging Simulations Through Active Learning

This paper presents an integrated, multi-disciplinary framework for Active Learning (AL) enabled scale-bridging approach. This approach is demonstrated through two exemplar applications, Inertial Confinement Fusion (ICF) and nanotransport in shale. The framework relies on a coarse-scale model, such as Lattice Boltzmann Method (LBM) or kinetic multi-ion Vlasov-Bhatnagar-Gross-Krook (Multi-BGK) model, which is divided into subdomains that are partitioned across nodes. The Active Learner Neural Network (NN) either generates an approximation from existing results or performs a new Fine-Scale Molecular Dynamics (MD) computation, depending on the estimate of the Uncertainty Quantifier (UQ) NN. The computational realization of these algorithms requires several aspects of next-generation architecture/algorithm design, including the reliance on an asynchronous task-based computing paradigm and the use of tensor cores for the NN training and inference.