Neural Network Domain Adaptation for Addressing the Generator-Dependence Problem in Impact Parameter Estimation

Abstract This work addresses the estimation of the impact parameter in heavy-ion collisions using simulated data from microchannel plate (MCP) detectors planned for future NICA experiments [1]. Neural networks can reconstruct the impact parameter accurately, but their performance depends strongly on the chosen event generator. We compared several approaches: principal component analysis, autoencoders, and naive mixed-dataset training did not yield generator-independent features. We then applied domain-adaptation methods, including domain-adversarial and deep reconstruction neural networks (DRNN). DRNN delivered the best performance, reducing generator bias while preserving sensitivity to the impact parameter, especially for central collisions. This may be a promising way toward generalized algorithms that can be reliably applied to forthcoming experimental data.