Electron-nucleus cross sections from transfer learning
Date:
Abstract: We present a deep learning approach to modeling inclusive electron–nucleus scattering cross sections using transfer learning, demonstrating how neural networks can effectively learn nuclear physical properties from limited experimental data. Initially trained on high-statistics electron–carbon scattering data, our model captures latent features that are transferable to other nuclei, such as lithium, oxygen, aluminum, calcium, and iron. This approach enables accurate predictions even in data-sparse regimes, revealing a robust internal representation of nuclear responses across different targets. Our results highlight the potential of representation learning to extract universal physical patterns and support data-driven modeling in nuclear and particle physics. The study illustrates how abstract, learned features can encapsulate domain knowledge, enabling generalization beyond the training distribution and reducing reliance on traditional theoretical models. To probe the method’s limitations, we applied it to the helium-3 target. The approach remained effective, although it required more extensive re-optimization. Finally, I will discuss the implications of this method for modeling neutrino-nucleus interactions. This talk is based mainly on the paper arXiv:2408.09936.