Neural network extraction of chromo-electric and chromo-magnetic gluon masses

We present a neural network-based quasi-particle model to separate the contributions of chromo-electric and chromo-magnetic gluons. Using dual residual networks, we extract temperature-dependent masses from SU(3) lattice thermodynamic data of pressure and trace anomaly. After incorporating physics regularizations, the trained models reproduce lattice results with high accuracy over $T\mathrm{/}{T}_c\in [1,10]$, capturing both the crossover behavior near $T_c$ and linear scaling at high temperatures. The extracted masses exhibit a physically reasonable behavior: they decrease sharply around $T_c$ and increase linearly thereafter. We find significant differences between thermal and screening masses near $T_c$, reflecting non-perturbative dynamics, while they converge at $T\gtrsim 2T_c$.