Physics Informed Neural Networks (PINNs) for neutronic equations

Sammanfattning

Artificial Intelligence (AI), and more specifically Physics-Informed Neural Networks (PINNs), is playing an increasingly pivotal role in modern scientific and industrial applications, driving innovation across diverse fields. This thesis, "Physics Informed Neural Networks (PINNs) for neutronic equations," explores the potential of AI-driven methods in neutronics, a critical area of nuclear engineering. In this paper, we analyze the feasibility and limitations of PINNs when applied to neutronic equations, which are characterized by their eigenvalue nature, multidimensional complexity, and multigroup energy formulations. The study also evaluates Data-Enabled PINN (DEPINN), an advanced framework, and its relevance to current industrial applications. By addressing feasibility, accuracy, and practical constraints, this research aims to explore new opportunities for leveraging AI in core physics modeling. The results demonstrate that while the PINN framework yields satisfactory outcomes for solving the two-group diffusion equation on simple geometries, such as multiple fuel assemblies, it struggles to converge to a physically accurate solution for more complex systems, such as full-core nuclear reactor models. To address this limitation, sensor data or known flux points are integrated into the model, transitioning to a DEPINN framework. The DEPINN approach proves promising, achieving high-quality flux maps and accurate multiplication factors (keff) when provided with data representative of operational sensors (1,300 MW reactor). These capabilities highlight its potential for industrial applications.