Making ICF Models More Predictive: Combining Simulations, Experiments and Expert Knowledge using Machine Learning and Bayesian Statistics

Understanding current inertial confinement fusion experiments, and designing future ones, is reliant on computer simulations which aim to predict implosion performance as a function of experimental design parameters. State of the art simulations, while highly accurate, are not predictive in the sense that post-shot tuning is required to match a given observation. The traditional approach to solving this problem, locally calibrating models by adding extra degradation mechanisms has been successful in explaining NIF experiments [1]; however, it is difficult to justify the extrapolation of locally calibrated models to new experiments or scales. Our ICF data is particularly challenging in this regard due to the extremely sparse nature of existing experiments, the highly nonlinear dependence of implosion performance on engineering parameters, and the difficulties in diagnosing integrated experiments.

In this talk, we will present a new predictive model that combines information from large-scale simulation studies and experiments [2]. We use a Bayesian input space calibration approach that combines a diverse set of experimental observables and accounts for all sources of uncertainty. The model is applied to a series of NIF ‘BigFoot’ shots [3] and is used to improve our understanding of current ICF experiments and their underlying physics, as well as to suggest the most fruitful paths to high yield.

LLNL-ABS-753911

[1] Kritcher et al., Physics of Plasmas 25(5), 056309 (2018)

[2] Gaffney et al., Statistical Analysis and Data Mining (2018) submitted

[3] Baker et al., Physical Review Letters (2018) submitted;

Casey et al., Physics of Plasmas 20(5), 056318 (2018)