Speaker
Description
Protons in the CLAS12 Forward Detector lose energy while passing through detector material, mainly through ionization, which shifts the reconstructed momentum away from the true value. This effect worsens momentum resolution and introduces biases into reconstructed kinematics, making energy loss corrections necessary for precision physics analysis.
We study machine-learning-based approaches to correct for this effect using simulated proton samples. In particular, we test supervised regression methods that learn either the true proton momentum or the momentum correction $\Delta p = p_{\mathrm{rec}} - p_{\mathrm{gen}}$ from reconstructed observables such as proton momentum and polar angle. The ML tools include dense feed-forward neural networks and gradient-boosted decision trees. We also compare different input feature sets and evaluation metrics to determine the most effective correction strategy. Model performance is assessed by comparing predicted and true proton kinematics and by studying residual distributions as functions of momentum to evaluate the stability and accuracy of the corrections in CLAS12.