At the US Lawrence Livermore National Laboratory (LLNL) global security is one of the core missions. When people's safety is at stake determining accurate error bounds on predictions and uncertainty quantification (UQ) is of vital importance. LLNL counts with the biggest supercomputers on the planet, along with the highest fidelity physics models. However, a costly single prediction is not enough if uncertainties are to be considered. For that purpose, many simulations are needed but the cost of performing millions of high-fidelity simulations is intractable.

Over the last decade, machine learning (ML) approaches have been increasingly relied upon for predicting responses of interest and UQ. The strengths of ML approaches are their speed and the ever-increasing ability to predict scenarios where experimental or simulation data is not available. ML approaches have been applied to accelerate predictions for applications where accurate, fast responses are a priority. In this presentation, I will show examples of ML applications and how their UQ is performed. The main presentation will cover the prediction and UQ of wind-driven patterns. K-nearest neighbors and deep-autoencoders have been used for prediction and their UQ is performed using metrics such as Jaccard index, fraction of data, fractional bias. The remaining part of the presentation will cover UQ in other areas: earthquake vs. explosion classification, source localization (inverse problem), and seismic-phase travel times.