The Best Possible Prediction: statistical inference and uncertainty quantification in predictions for ICF experiments

To make the best possible prediction, we must combine models and data optimally. As an example, we apply a method for inference on large data sets to the problem of ``predicting'' the unknown results, with uncertainties, of 16 direct-drive implosion experiments, using glass and plastic capsules, shot at OMEGA. The method uses the GPM/SA (Gaussian Process Models for Simulation Analysis) code to construct an emulator, based on a rad-hydro code and constrained by data from 22 other ICF implosions carried out under different conditions from the unknown set. Comparing the extrapolative ``predictions'' to the actual observations lets us evaluate the validity of various assumptions and the reliability of the predictions and uncertainty bounds [Osthus et al., SIAM/ASA J. Uncert. Quant. 7, 604 (2019)]. The predictions turn out to be quite reliable: 94{\%} of the predictions agreed with the actual observations to within the 95{\%} uncertainty bounds. This approach will likely also be useful for model calibration and validation, hypothesis testing, and experiment design.