Learning and correcting non-Gaussian model errors

All discretized numerical models contain modelling errors - this reality is amplified when reduced-order are used. The ability to accurately approximate informs statistics on model confidence and improves quantitative results from frameworks using in prediction, tomography, signal processing. Further this, the compensation of highly nonlinear non-Gaussian errors, arising many ill-conditioned systems aiming capture complex physics, a historically difficult task. In work, we address challenge by proposing neural network approach capable approximating compensating for such augmented direct inverse problems. viability demonstrated simulated experimental data differing physical