Improving Initialization of ASTEC's Thermal-Hydraulic Solver with Machine Learning-Based Methods for Enhanced Convergence in Severe Accident Simulations

Severe Accident (SA) simulation codes like ASTEC (Accident Source Term Evaluation Code) are essential for predicting nuclear reactor behavior under SA conditions. However, their high computational demands, particularly in thermal-hydraulic simulations (e.g., the CESAR module in ASTEC), can constrain their effectiveness. A significant portion of CESAR’s computational load arises from solving non-linear Partial Differential Equations (PDEs) at each timestep using the Newton-Raphson (NR) iteration method. While NR convergence depends on having initial guesses close to the final solution, the current ASTEC implementation relies solely on values from previous converged states, without predictive insights. This paper introduces a hybrid approach aimed at enhancing CESAR's NR solver through Machine Learning (ML) models trained to learn temporal dynamics via reconstruction, laying the groundwork for future predictive initialization. Drawing on recent advancements, this approach explores using ML based surrogate models to learn the intricate, non-linear relationships within transient conditions, aiming to reduce the number of iterations needed for convergence and potentially allow longer timestep intervals without sacrificing accuracy. The choice of surrogate model remains adaptable, seeking to balance predictive accuracy and computational efficiency within CESAR’s frequent initialization routines. Preliminary results show that an ML augmented approach can reduce ASTEC's computation time, suggesting promising implications for broader applications in thermal-hydraulic simulations in nuclear safety assessments. Future work may also involve developing an ensemble of surrogate models, complemented by a classifier that dynamically selects the most suitable initialization based on the reactor state and accident phase, optimizing the solver’s performance across varying scenarios.