A Robust Learning Methodology for Uncertainty-Aware Scientific Machine Learning Models

Robust learning is an important issue in Scientific Machine Learning (SciML). There are several works the literature addressing this topic. However, there increasing demand for methods that can simultaneously consider all different uncertainty components involved SciML model identification. Hence, work proposes a comprehensive methodology evaluation of also considers possible sources uncertainties identification process. The considered proposed method absence theory, causal models, sensitivity to data corruption or imperfection, and computational effort. Therefore, it provide overall strategy uncertainty-aware models field. validated through case study developing soft sensor polymerization reactor. first step build nonlinear parameter probability distribution (PDF) by Bayesian inference. second obtain machine Monte Carlo simulations. In step, PDF with 30,000 samples built. evaluated sampling 10,000 values simulation. results demonstrate identified sensors robust uncertainties, corroborating consistency approach.