Extraction of the microscopic properties of quasi-particles using deep neural networks

Abstract: We use deep neural networks (DNN) to obtain the microscopic characteristics of partons in terms of dynamical degrees of freedom on the basis of an off-shell quasiparticle description. We aim to infer masses and widths of quasi-gluons, up/down, and strange quarks using constraints on the macroscopic thermodynamic observables obtained by the first-principles calculations lattice QCD. In this work, we use 3 independent dimensionless thermodynamic observables from lQCD for minimization. First, we train our DNN using the DQPM (Dynamical QuasiParticle Model) Ansatz for the masses and widths. Furthermore, we use the DNN capabilities to generalize this Ansatz, to evaluate which quasiparticle characteristics are desirable to describe different thermodynamic functions simultaneously. To evaluate consistently the microscopic properties obtained by the DNN in the case of off-shell quarks and gluons, we compute transport coefficients using the spectral function within Kubo-Zubarev formalism in different setups. In particular, we make a comprehensive comparison in the case of the dimensionless ratios of shear viscosity over entropy density $\eta/s$ and electric conductivity over temperature $\sigma_Q/T$, which provide additional constraints for the parameter generalization of the considered models.