Speaker
Description
Relativistic heavy-ion collisions serve as a primary tool for investigating the fundamental properties of matter under extreme conditions. The theoretical modeling of these interactions relies on various computational models whose predictive power often fluctuates across different kinematic ranges and physical observables. Furthermore, the underlying complex phenomenological chains are computationally intensive, often creating a bottleneck for statistic intensive physics analyses.
In this work, we present a novel approach to unify the strengths of different simulation engines through a Generative Adversarial Network (GAN) featuring a multi-discriminator architecture. This architecture allows the generator to learn from a heterogeneous training dataset comprising multiple simulation engines, effectively bridging the performance gaps across various parameter spaces at the single-event level. We demonstrate that this specialized learning improves the fidelity of generated physical distributions while reducing the computational footprint compared to conventional methods. The performance of the model is evaluated against key experimental observables, showing its potential for large-scale data analysis in future high-energy physics experiments.