Gianluca Gregori

Abstract:

Computer simulations are invaluable tools for scientific discovery. However, accurate simulations are often slow to execute, which limits their applicability to extensive parameter exploration, large-scale data analysis, and uncertainty quantification. A promising route to accelerate simulations by building fast emulators with machine learning requires large training datasets, which can be prohibitively expensive to obtain with slow simulations. Here we present a method based on neural architecture search to build accurate emulators even with a limited number of training data. The method successfully accelerates simulations by up to 2 billion times in 10 scientific cases including astrophysics, climate science, biogeochemistry, high energy density physics, fusion energy, and seismology, using the same super-architecture, algorithm, and hyperparameters. Our approach also inherently provides emulator uncertainty estimation, adding further confidence in their use. We anticipate this work will accelerate research involving expensive simulations, allow more extensive parameters exploration, and enable new, previously unfeasible computational discovery.

Abstract:

Our understanding of physical systems often depends on our ability to match complex computational modeling with the measured experimental outcomes. However, simulations with large parameter spaces suffer from inverse problem instabilities, where similar simulated outputs can map back to very different sets of input parameters. While of fundamental importance, such instabilities are seldom resolved due to the intractably large number of simulations required to comprehensively explore parameter space. Here, we show how Bayesian inference can be used to address inverse problem instabilities in the interpretation of x-ray emission spectroscopy and inelastic x-ray scattering diagnostics. We find that the extraction of information from measurements on the basis of agreement with simulations alone is unreliable and leads to a significant underestimation of uncertainties. We describe how to statistically quantify the effect of unstable inverse models and describe an approach to experimental design that mitigates its impact.