Using Neural Networks to Predict Pin Powers in Reflective PWR Fuel Assemblies with Varying Pin Size

The use of neural networks to predict high-fidelity neutronics features is becoming an increasingly attractive area of investigation, as a way to reduce the computational resources needed for simulations while maintaining the high resolution of the latent simulations. Previous work provided a novel network architecture, LatticeNet, as an approach to use neural networks to predict high-resolution pin power predictions equivalent to what would be produced by a high-fidelity code without significant computational cost. This paper further tests this approach by applying it in scenarios with varying fuel pin sizes, and shows that it can be successfully used to achieve high accuracy in predictions, even in regions which the training data did not explicitly represent.