ANICCA fuel cycle irradiation models: a machine learner predictor as a function of initial fuel composition

Machine Learning and Deep Learning techniques are gaining popularity due to the numerous applications they serve, being one of this applications the use of neural networks for predicting values from input data. This implies the possibility of replacing relatively complex or computationally expensive mathematical models with these neural networks.
Therefore, the present work, developed at the Belgian Nuclear Research Centre (SCK-CEN), has been focused on the use of these neural networks for the substitution of more complex processes. In this case, the substitution of the Chebyshev Rational Approximation Method (CRAM) was achieved. CRAM is used in the in-house Python-coded ANICCA nuclear fuel cycle simulator to solve the Bateman depletion equations to calculate the final isotopic inventory after an irradiation process.
The main purpose of this work was to replace the ANICCA irradiation module that hosted the CRAM with a model based on these neural networks that would have as input the initial fuel enrichment (or plutonium content if MOX) and discharge burnup, since such information is easily retrievable from within the current ANICCA code.
For this purpose, two neural networks were generated, one for cycles involving conventional Uranium fuel, and one for those using MOX fuel. These networks were trained from two datasets generated from depletion calculations performed in SERPENT2. These datasets encompassed many of the normal enrichment values for UOX and plutonium content for MOX, as well as numerous burnup steps covering all conventional discharge burnup values.
After obtaining the data and optimizing the hyperparameters and then training the neural networks in the Keras API for Deep Learning, the neural networks were able to predict in the final inventory the mass density of 73 isotopes relevant for spent nuclear fuel characterization and fuel cycle analysis. The observed deviations between SERPENT2 and the output of the networks were less than 10% for 80% of the tested values. However, some greater deviations were found for heavy nuclides such as Californium and Curium.
The neural network models were then saved and deployed in ANICCA's irradiation module to test the performance of these models by means of a benchmark in scenario studies. These benchmarks resulted in a successful integration of the modifications, with deviations of less than 10% for the final inventories of Uranium, Plutonium, Fission Products and Transuranium elements. However, a slightly higher deviation was observed for the minor actinides. At the same time, computation time was reduced by a solid 50% for complex scenarios.
However, the rigidity of this approach was highlighted since the MOX model can only infer the inventory for fuels with a single plutonium vector and can only predict as many isotopes as the network has been trained for. To address these limitations, new neural networks can be trained, improving the accuracy of predictions by adding more inputs, such as the cooling time to deal with short-lived isotopes, and the weight percentage of different isotopes in the fresh MOX fuel to predict several MOX compositions. Finally, computational time could also be enhanced by applying different and more advanced and specific deployment and coding techniques for the neural networks design.