Machine Learning for Thermal Hydraulics

Re-running CFD for every change of a design variable fills a design review with waiting. Decomposing the turbulent cross flow in a helical coil steam generator with Proper Orthogonal Decomposition (POD), which splits a flow field into a few representative shapes, keeping only the dominant modes, and learning how those modes evolve with an LSTM gives the same prediction far faster. This reduced-order model work was presented at NURETH-21 and published in a special issue of Nuclear Engineering and Technology.

The same idea serves measurement. Two bubble silhouettes taken by two cameras at right angles carry no depth, yet a trained 3D convolutional neural network recovers the bubble's volume from them voxel by voxel. The method is also the subject of a patent application.

We also work with physics-informed neural networks (PINNs), which put the governing equations directly into the loss function. Applying an adaptive dynamic adjustment to neutron transport in simple geometries, we study how to keep a network from drifting away from the physics when data is scarce.

In every case it is the quality of the training data that decides the outcome. So building high-fidelity CFD datasets for SMR thermal hydraulics is itself a line of work here. Unvalidated data does not go into training.

And we look past that. For now each model learns one flow, one geometry. What we actually want is a model that has learned how a reactor behaves as a whole, one you can hand a changed condition and let it roll forward on its own. A world model, in other words. It is a long way off, but the reduced-order models and physics-constrained learning we are building now are the footholds on the way there.