Experimental and numerical investigation on debris bed quenching and development of an artificial neural network for quick quenching estimation

A debris bed can form in the reactor cavity during a beyond-design-basis accident of light water reactors with core degradation and RPV failure. With insufficient cooling water, the debris bed can melt, interacting with the concrete underneath and generating non-condensible gases at the bottom, affecting coolability for re-flooding the superheated particle bed. Thermal-hydraulic system codes like ATHLET can, in principle, consider the impact of an additional gas flow on the quenching process. However, there is still a need for experimental validation of the respective models and the validation of the corresponding simulation results. A specific extension to the existing experimental database is needed for the model validation of COCOMO, which is implemented in ATHLET. Since detailed simulation codes like COCOMO can require long computation times, there is a need for fast-running models for probabilistic risk analysis. Artificial neural networks can be utilised for quick estimations of the quenching process. Experimental results are presented for the top-flooding quenching behaviour of a monodisperse particle bed with the influence of additional non-condensable gas injection conducted at the FLOAT test facility. Consecutive numerical simulations are carried out for model validation. Since multi-dimensional simulations are computationally expensive, an artificial neural network is developed for quick estimation of the quenching process. COCOMO simulations capture the basic phenomenology of top-flooding quenching but deviate due to modelled 3D effects and underestimate the influence of NCG injection on quenching time. The developed ANN similarly underestimates this effect but shows promise for quick quenching estimation. • Top-flooding tests with non-condensible gases have been carried out. • Non-condensible gas injection significantly increased quenching time. • COCOMO can describe the phenomenology in simplified two dimensional simulations. • Artificial neural networks are trained to predict the quench front propagation. • The predictions generally are in good agreement with COCOMO simulations.