Speakers
Description
In many engineering studies, computer codes are increasingly used to understand, model, and predict physical phenomena. However, the numerical models underlying these codes can be computationally expensive, severely limiting the number of simulations that can be performed. A widely adopted solution consists in replacing the expensive computer code with a computationally efficient mathematical approximation, referred to as a surrogate (or meta-model). Surrogates may rely on a broad range of supervised learning techniques, including polynomial regression, Gaussian processes, random forests, and neural networks. Trained on a set of numerical simulations, they should accurately reproduce the code outputs over the input domain of interest while maintaining strong predictive performance at unseen points.
This talk addresses the challenges involved in validating surrogates intended for use in safety-critical systems, whose failure could have catastrophic consequences, with a particular focus on nuclear safety applications. The work was initially conducted by a working group dedicated to assessing the reliability of thermal-hydraulic passive systems.
We first examine the fundamental differences between traditional physics-based numerical models and data-driven surrogates constructed using machine-learning algorithms. Rigorous validation of these tools is essential to establish their reliability and suitability for critical applications. We also discuss connections with recent developments in trustworthy artificial intelligence.
Drawing inspiration from the standardized Verification, Validation, and Uncertainty Quantification process applied to scientific computing tools in nuclear safety studies, we propose a methodological framework structured around four confidence dimensions: conformity, robustness, explainability, and transparency. Each dimension must be assessed using appropriately selected methods and criteria. A case study concerning the reliability of thermal-hydraulic passive systems illustrates the relevance and practical implementation of this framework.