This study is conducted in the frame of the EU-funded project ASSAS (Artificial intelligence for Simulation of Severe AccidentS), which aims at developing a basic-principles severe accident simulator for a generic PWR-1300 MW and at evaluating the possibility to replace some ASTEC models with machine-learning surrogate models. This paper presents the preliminary stage of the surrogate model development, focusing mostly on the generation of training datasets necessary for the machine learning process, and addresses essentially the CESAR module, which handles thermal-hydraulics in the primary and secondary circuits. To achieve this, a base case scenario involving a Station Blackout (SBO) without any safety measures until Reactor Pressure Vessel (RPV) failure is considered. From this base case, various calculations are performed by introducing safety measures (e.g. low-pressure injections) at different times. This safety measures sampling is applied within the time range between the onset of the primary pressure drop (enabling low-pressure injections) and the RPV failure initially predicted in the base case.

Two approaches are employed. The first involves conducting a batch of 22 runs, varying one or two criteria in each run, and saving most relevant ASTEC variables at each time step (especially state variables from CESAR cells, but also system signals such as pump actuation). This should generate a suitable dataset for the development of time-stepping surrogate models, that are expected to compute each time-step, like ASTEC does, and also account for user’s decisions during the accident calculation interactively.

Alternatively, another approach would consist of developing a meta-model that predicts key outcomes of the SBO - such as the delayed RPV failure time or final hydrogen production - with respect to one criteria (to simplify the problem): the safety injection time. However, due to the intricate mechanisms leading to RPV failure, training the neural network requires a dataset with a larger amount of samples, i.e. a finer variation of the injection time; here, a batch of 110 runs. Unlike the previous approach, preserving results at every time step is not as necessary. This last strategy is quite independent of the ASSAS project, as this global approach cannot suit the requirements of an online simulator; the purpose here is more exploratory. The last part of the paper proposes machine-learning strategies and discusses the optimal utilization of the generated training datasets.