Skip to main content
Back to timeline
arXivSource publication:

Neural networks learn temperature, field and scale dependence of the effective potential without a large-N crutch

Synopsis

Solving functional renormalization group flow equations now uses physics-informed neural networks without the large-N approximation of earlier attempts and without any grid discretization. The approach combines small-field and large-field expansions into a single ansatz whose coefficients are predicted by a network, approximating the effective potential across the whole field range. In the local potential approximation of a finite-temperature O(N)-symmetric scalar theory, raising the hidden size to 128 makes the network's temperature dependence of the order parameter practically indistinguishable from grid-based results.

AI-generated editorial illustration: Utilizing Neural Networks for Solving Functional Renormalization Group Flow Equations

Interpretation

Functional renormalization group flow equations can now be solved by a grid-free neural network ansatz that learns the effective potential's dependence on temperature, field and renormalization scale at once. Earlier physics-informed neural network schemes relied on the large-N approximation, using the analytic large-N solution as a baseline and training the network only on finite-N corrections, a decomposition not guaranteed to exist for more complex multi-component scalar models. In the local potential approximation of a finite-temperature O(N)-symmetric scalar theory, raising the hidden size to 128 makes the network's temperature dependence of the order parameter practically indistinguishable from grid-based results, which is the main result.

The network does not fit the effective potential directly; it serves as a building block of an ansatz that joins large-field and small-field expansions to cover the entire field range. The gradual flattening of the effective potential toward convexity in the small-field region is the numerically hardest part, and a large-field form alone cannot capture it accurately. The ansatz went through three modifications, progressively adding a power-series form, network-predicted transition scales and the renormalization scale as an input, and the final form was tested in both fixed-temperature and varying-temperature training.

Derivatives come analytically from the ansatz instead of finite-difference approximations, so no numerical instability from grid discretization is introduced. The earlier scheme obtained derivatives of the potential through finite differences, which tend to become unstable for multi-variable functions. Training runs on the local potential approximation flow equation, with the loss summing partial differential equation residues over randomly drawn temperature, field and scale triplets, optimized with ADAM and L-BFGS.

Perspective

The results concern the local potential approximation of a finite-temperature O(N)-symmetric scalar theory, with grid-based solutions as the comparison for temperature, field and scale dependence. The authors propose extending the method to systems whose effective potential depends on multiple field variables, to scalar theories with the corresponding symmetries, and to testing its applicability to gauge theories.

Varying-temperature training reaches a terminal loss more than an order of magnitude higher than fixed-temperature training at the same network size, so precision is partly sacrificed; the authors note that larger hidden sizes are needed to represent the temperature dependence accurately, and the parameter budget required to lower the infrared cutoff while keeping accuracy is not investigated in detail.