This neural network architecture, LatticeNet, uses computer vision, modular neural networks, and deep learning approaches to optimize the fuel rod configuration in a nuclear reactor stack in the time it takes a low-fidelity application to do one run. Available applications in machine learning are either high-fidelity, requiring hundreds of node-hours to get high performance, or low-fidelity, requiring only seconds on common computing hardware. Research needs the best attributes of both high and low fidelity.
Researchers at the University of Florida have developed a neural network architecture that combines both methods of machine learning to optimize configurations of nuclear reaction rods. Entities replacing rods and building new facilities can use the calculations to maximize output and fuel rod life.
Optimize configuration of nuclear reactor fuel rods
In neural networks, each transformation in these layers of transformations is not a single continuous-valued transformation but rather a series of independent vector-to-scalar transformations of the input provided by the previous transformation layer (or the network input). It is demonstrated that LatticeNet, when tuned using this methodology, can effectively predict the normalized pin powers with less than 0.005 absolute error per-pin in most cases, even when including common burnable poison types.
