Speaker
Description
Generative models have demonstrated potential in accelerating Monte Carlo sampling for physical simulations, yet their capability to handle strongly correlated point cloud data, such as atomic nuclei structures, remains unclear, which directly affects their reliability in nuclear physics simulations. Taking the oxygen-16 nucleus as an example, this paper systematically evaluates the performance of generative models—including normalizing flows, diffusion models, and flow matching(FM)—in learning nuclear structures. By analyzing single-nucleon distributions and two-nucleon correlation functions, we find that traditional models struggle to accurately reproduce short-range correlations between nucleons, whereas Transformer-based flow matching models effectively learn nucleon correlation patterns. The study indicates that the choice of neural network architecture is crucial: the attention mechanism in Transformers provides a key pathway for modeling two-body and three-body correlations. Using the optimized generative model, we achieved a sampling speed improvement of approximately 250 times for nucleon positions. This offers a more efficient solution for Monte Carlo simulations of high-energy nuclear collisions and holds promise for advancing subsequent Bayesian analysis and refined quantitative estimation.
