Transfer Learning for Neutrino Scattering: Domain Adaptation with GANs

Abstract

Transfer learning (TL) is used to extrapolate the physics information encoded in a Generative Adversarial Network (GAN) trained on synthetic neutrino-carbon inclusive scattering data to related processes such as neutrino-argon and antineutrino-carbon interactions. We investigate how much of the underlying lepton-nucleus dynamics is shared across different targets and processes. We also assess the effectiveness of TL when training data is obtained from a different neutrino-nucleus interaction model. Our results show that TL not only reproduces key features of lepton kinematics, including the quasielastic and $Δ$-resonance peaks, but also significantly outperforms generative models trained from scratch. Using data sets of 10,000 and 100,000 events, we find that TL maintains high accuracy even with limited statistics. Our findings demonstrate that TL provides a well-motivated and efficient framework for modeling (anti)neutrino-nucleus interactions and for constructing next-generation neutrino-scattering event generators, particularly valuable when experimental data are sparse.

Figures (22)

• Figure 1: Test data samples $(E'_\mu,\theta')$ distributions for $E_\nu = 500$ MeV. On the top left, we show the original MC $\nu_\mu$-carbon scattering; on the top right, the $\nu_\mu$-argon scattering; on the bottom left, the $\bar{\nu}_\mu$-carbon scattering; and on the bottom right, the distributions of the alternative $\nu_\mu$-carbon samples.
• Figure 2: Generator architecture. The blue block encloses the portion of the network that remains frozen when training with TL. The output of the network is given by the proxy vector Eq. \ref{['Eq:proxy']}. Similar to Ref. bonilla2025generativeadversarialneuralnetworks, the skip connections are applied but not shown in the graph above.
• Figure 3: Discriminator architecture. The blue block encloses the portion of the network that remains frozen when training with TL. The network takes for the input $\mathbf{v}$ (the proxy variable vector defined by Eq. \ref{['Eq:proxy']}) and proxy neutrino energy $E'_\nu$. Similar to Ref. bonilla2025generativeadversarialneuralnetworks, the skip connections are applied but not shown in the graph above.
• Figure 4: Training data $450 < E_\nu < 550$ MeV sub-samples $(E'_\mu,\theta')$ distributions. On the top, we have the sub-sample from a training sample of $100$k events, while on the bottom is the sub-sample from a sample of $10$k events.
• Figure 5: Dependence of the MAP-3D metric on the number of optimization epochs for the $\nu_\mu$–argon models. The left and right columns correspond to optimizations using 10,000 and 100,000 events, respectively. The top and bottom rows show results for models trained from scratch and with transfer learning (TL), respectively. Red dashed vertical lines mark the epochs of the selected models.