EINN2025

Modeling lepton-nucleus scattering cross sections using transfer learning technique

Not scheduled

20m

Talk in workshop 2: "AI & ML in nuclear science: starting with design, optimization, and operation of the machine and detectors, to data analysis"

Speaker

Beata Kowal (University of Wrocław)

Description

The neural network framework allows us to obtain empirical fits to electron scattering cross sections on carbon over a wide kinematic region. Transfer learning makes it possible to adapt a deep neural network trained on one type of data to new problems. We apply transfer learning to derive a new model from a previously obtained set of neural networks trained on electron-carbon cross-section data. Using the bootstrap method, we retrain this set of networks on cross-section data for electron interactions with other nuclear targets. This procedure yields fits and their uncertainties for helium, lithium, oxygen, aluminum, calcium, and iron.

In our analysis, the loss function is defined by the χ2, which incorporates both point-to-point and normalization uncertainties for each independent dataset. Since electron–nucleus and neutrino–nucleus interactions share many similarities, our technique has the potential to significantly improve the understanding of neutrino interactions with nuclei, which is crucial for neutrino oscillation experiments such as DUNE and Hyper-Kamiokande.

The presentation is based on the papers Phys.Rev.C 110 (2024) 2, 025501; Phys. Rev. Lett. 135, 052502; arXiv:2508.00996