# Logical Reasoning for Revealing the Critical Temperature through Deep Learning of Configuration Ensemble of Statistical Systems

## TL;DR

This paper demonstrates that deep learning models can identify the critical temperature of phase transitions in statistical systems by analyzing bias parameters, which encode free energy, rather than weight parameters.

## Contribution

It establishes a theoretical link between deep learning parameters and physical quantities, revealing bias parameters encode free energy and can detect critical points.

## Key findings

- Bias parameters record free energy as a function of temperature.
- Deep learning reveals critical temperature via second difference of biases.
- Weight parameters do not serve as order parameters, contrary to previous claims.

## Abstract

Recently, there have been many works on the deep learning of statistical ensembles to determine the critical temperature of a possible phase transition. We analyze the detailed structure of an optimized deep learning machine and prove the basic equalities among the optimized machine parameters and the physical quantities of the statistical system. According to these equalities, we conclude that the bias parameters of the final full connection layer record the free energy of the statistical system as a function of temperature. We confirm these equalities in one- and two-dimensional Ising spin models and actually demonstrate that the deep learning machine reveals the critical temperature of the phase transition through the second difference of bias parameters, which is equivalent to the specific heat. Our results disprove the previous works claiming that the weight parameters of the full connection might play a role of the order parameter such as the spin expectation.