Inclusive electron-nucleus cross section models from domain adaptation

Developing accurate theoretical models for electron-nucleus scattering is essential for benchmarking neutrino-nucleus interaction models used in oscillation experiments. This paper investigates whether transfer learning (TL) can effectively adapt a high-quality deep neural network (DNN) model, originally trained on electron-carbon data, to describe inclusive electron scattering on other nuclear targets.

The authors utilize an ensemble of 50 DNNs pretrained on carbon data as a baseline. For each target (3He, 6Li, 16O, 27Al, 40Ca, and 56Fe), they perform fine-tuning by retraining a subset of the network layers using target-specific data. They systematically evaluate the model performance by varying the number of fine-tuned layers (layer-wise analysis), the fraction of training data (70% vs. 10%), and the kinematic overlap between the source (carbon) and target domains. The models are validated against experimental data and compared with the phenomenological F1F2 model.

The study confirms that the carbon-pretrained model captures universal kinematic features relevant to a wide range of nuclei. Oxygen requires only shallow adaptation, while heavier or more complex nuclei like iron and calcium require deeper fine-tuning. The transfer learning approach is shown to be robust for most targets, significantly outperforming the untuned baseline. However, the quality of the adapted model is sensitive to the available data; lithium, with its sparse dataset, is the least robust case, and aluminum's performance is heavily influenced by a small subset of high-precision measurements.

This work provides a scalable, data-driven method for constructing inclusive electron-nucleus cross-section models. By leveraging existing carbon data, this approach reduces the need for large, target-specific datasets, making it a powerful tool for nuclei relevant to future neutrino experiments, such as argon, where data may be limited. It also offers a systematic way to quantify model uncertainty and evaluate the limits of transferability in nuclear physics.