Deep Learning Nuclear Cross Sections: Extrapolating to the Dripline with GNN-VAE Frameworks.

Machine learning models have recently shown great promise in capturing the underlying trends and features of nuclear cross sections across the chart of nuclides [1]. Building on a framework that combines variational autoencoders (VAEs) and graph neural networks (GNNs) to encode nuclear data and learn intrinsic correlations in the latent space, we extend this approach to include neutron-induced fission ((n,f)) channels and evaluated nuclear data libraries . These additions aim to improve the representational power and predictive accuracy of the network, particularly in regions where theoretical models remain uncertain.

We assess the impact of these extensions on cross section prediction accuracy and the structure of the learned latent space. Using the augmented framework, we generate extrapolated cross section predictions for (n,n), (n,n’), (n,g) and (n,f) reactions across the nuclear chart, including nuclei approaching the neutron dripline. This work represents a step toward unifying evaluated nuclear data libraries under a common deep-learning-based representation, with potential applications in nuclear technology, astrophysics and data evaluation.

References:

[1] Mintra, S., et. al., “Learning nuclear cross sections across the chart of nuclides with graph neural networks”, (2025). https://arxiv.org/abs/2404.02332