Skip to main content

About

Use of Artificial Intelligence:
data-driven reduced-order modelling

To maximise the chances of success, different options will be explored in parallel to improve the performances of ASTEC. Besides classical optimisation, machine-learning models may achieve drastically higher speed-up factors.

The different techniques to be explored are the following:

  • Linear and non-linear dimension reduction,
  • Time-stepping approaches to enable the hybridation of machine-learning and physics-based models,
  • Time-series prediction for global models replacing severe accident codes totally,
  • Interpolation between precalculated sequences.

The scope of the surrogate model may also vary:

  • A specific system (primary loops, GVs, reactor core…),
  • A model, a module or group of modules in ASTEC (thermal-hydraulics, core degradation…),
  • Complete code for a specific part of the simulation (post vessel-rupture phase…).

Machine-learning models will ultimately be used in the simulator, so they must be able to react to operator actions. It means that sequences cannot be calculated in advance, because the boundary conditions are not known at the start of calculations. 


The different options explored by partners will be described in a report on the modelling strategy that is expected in April 2024. The most promising models will be coupled with ASTEC or directly with the simulator. 

To reduce the dimensionality of the already very complex problem to solve, only best-estimate physical parameters will be used for the project, without considering their uncertainty.

Physics-informed models will be considered to respect some conservation laws natively, which will ultimately contribute to the trustworthiness of the developed algorithms.

A large quantity of data will be necessary to train the advanced architectures necessary for the project. It will be hosted by the Large Scale Data Facility operated by KIT.