Solving Functional Renormalization Group Equations with Neural Networks Preprint: RIKEN-iTHEMS-Report-26

Yang-yang Tan, Wei-jie Fu, Lianyi He, Lingxiao Wang · 2026

Read Solving Functional Renormalization Group Equations with Neural Networks Preprint: RIKEN-iTHEMS-Report-26 by Yang-yang Tan, Wei-jie Fu, Lianyi He, Lingxiao Wang (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-21151.txt.

Introduction

Tan and co-authors solve functional renormalization group equations with neural networks. The renormalization group tracks how an effective potential changes with scale, a core tool in nonperturbative field theory. They train networks to represent the potential's field derivative, with the flow equations built into the loss. No precomputed training data is needed. For the O(N) scalar theory, the method captures symmetric, broken, and critical regimes. Results match finite-difference and Galerkin methods well.

Catalog record

CREATORYang-yang Tan, Wei-jie Fu, Lianyi He, Lingxiao Wang
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE63847
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.21151

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Solving Functional Renormalization Group Equations with Neural Networks Preprint: RIKEN-iTHEMS-Report-26 Yang-yang Tan, Wei-jie Fu, Lianyi He, Lingxiao Wang Abstract We employ deep neural networks to represent the field derivative of the scale-dependent effective potential in the functional renormalization group (fRG) framework for nonperturbative quantum field theory. By embedding the fRG flow equations directly into the loss function, the network parameters are determined so as to provide a continuous and differentiable representation of the scale- and field-dependent effective potential without relying on precomputed training data. Focusing on the O ⁡ ( N ) O(N) scalar field theory within the local potential approximation at finite temperature, we demonstrate that this neural network representation accurately captures the renormalization group flow across symmetric, broken, and critical regimes. A key ingredient is a decomposition of the representation into an analytically known large- N N contribution and a learned finite- N N correction, which efficiently mitigates numerical stiffness associated with convexity restoration in the broken phase. The physics-driven solutions show excellent agreement with established finite-difference and discontinuous Galerkin methods. We further apply the same strategy to the Wilson-Fisher fixed point equation in three dimensions, illustrating that neural network representations provide a unified framework for both scale-dependent flows and fixed-point problems. Our results indicate that physics-driven deep learning offers a robust and flexible numerical tool for functional renormalization group studies. I Introduction Understanding the non-perturbative dynamics of strongly correlated quantum systems remains one of the central challenges in modern theoretical physics. From the confinement of quarks and the phase diagram of quantum chromodynamics (QCD) at finite temperature and density, to quantum phase transitions and critical phenomena in condensed matter systems, many of the most fascinating physical phenomena occur in regimes where conventional perturbative expansions fail. The functional renormalization group (fRG) Wetterich 1993 has emerged as a preeminent non-perturbative tool for addressing these challenges, implementing Kenneth Wilson’s renormalization group philosophy Wilson 1971a ; Wilson 1971b ; Wilson and Fisher 1972 ; Wilson and Kogut 1974 by systematically integrating out fluctuations scale by scale. The fRG is built upon the Wetterich equation Wetterich 1993 , an exact functional differential equation governing the scale dependence of the effective action \Gamma_{k}[\phi] . By introducing a momentum-dependent infrared regulator, the fRG smoothly interpolates between the classical action in the ultraviolet (UV) and the full quantum effective action in the infrared (IR). This approach has proven remarkably successful across diverse areas of physics, including the critical phenomena for strong correlated systems Berges et al. 2002 ; Canet and Chate 2007 ; Canet et al. 2011 ; Mesterházy et al. 2013 ; Bluhm et al. 2019 ; Roth and von Smekal 2023 ; Tan et al. 2025a ; Chen et al. 2025 ; Roth et al. 2025a ; Roth et al. 2025b , the QCD phase diagram Braun et al. 2016 ; Mitter et al. 2015 ; Rennecke 2015 ; Cyrol et al. 2016 ; Cyrol et al. 2018a ; Cyrol et al. 2018b ; Fu et al. 2020 ; Braun et al. 2020 ; Ihssen et al. 2024 ; Pawlowski et al. 2025 ; Tan et al. 2025b , and the quantum gravity. One can refer to relevant reviews, e.g., Berges et al. 2002 ; Pawlowski 2007 ; Braun 2012 ; Dupuis et al. 2021 ; Fu 2022 for more details. Despite its conceptual elegance and broad applicability, practical calculations of fRG usually have to face significant computational challenges. The Wetterich equation involves infinite-dimensional functional spaces that necessitate systematic truncation schemes such as the derivative expansion or the vertex expansion. Even after truncation, the finite sets of flow equations are still challenging to be solved numerically in many cases. In recent years, machine learning (ML) has emerged as a transformative tool across many branches of physics, offering new approaches to longstanding computational and theoretical challenges, especially in the QCD physics Boehnlein et al. 2022 ; Zhou et al. 2024 ; Aarts et al. 2025 . Within the context of quantum field theory, ML techniques have found applications in lattice QCD simulations Boyda et al. 2022 ; Cranmer et al. 2023 ; Aarts et al. 2025 , including generating gauge field configurations and accelerating Monte Carlo sampling. Diffusion models, connected to the stochastic quantization, have shown particular promise for lattice field theory simulations Wang et al. 2024 ; Wang et al. 2023 ; Zhu et al. 2024 ; Zhu et al. 2025 ; Aarts et al. 2026 . A particularly powerful paradigm is physics-informed Raissi et al. 2019 ; Karniadakis et al. 2021 or physics-driven learning Aarts et al. 2025 , where physical laws and symmetries are directly embedded into the network architecture or loss function, enabling solutions that respect the underlying physics by construction. In this work, we introduce a novel computational approach in which a neural network is used to represent the field derivative of the effective potential, with the fRG flow equations embedded directly into the loss function to determine the network parameters. This idea draws on the physics-informed neural network (PINN) Raissi et al. 2019 ; Karniadakis et al. 2021 , which incorporates governing differential equations, boundary conditions, and initial conditions simultaneously into the loss function and has demonstrated remarkable success across scientific computing, including fluid dynamics, solid mechanics, etc. Recently, PINNs have been applied to solve the functional renormalization group equations on lattices Yokota 2024 , or functional differential equations more generally Miyagawa and Yokota 2024 , and the Dyson-Schwinger equations Terin 2025 ; Carmo Terin 2025 . In contrast to the standard PINNs, our approach uses only the PDE residual of the fRG flow equation as the loss function, without imposing explicit boundary or initial conditions as additional constraints. In our case, we show that the neural network representations of the effective potential provide an effective approach for the fRG calculations. We demonstrate our approach with the O(N) scalar field theory within the local potential approximation (LPA), a paradigmatic model that captures the essential physics of spontaneous symmetry breaking and critical phenomena. The O(N) model encompasses a wide range of universality classes, including the Ising model ( N=1 ) for the liquid-gas and magnetic transitions, the XY model ( N=2 ) for the superfluidity, the Heisenberg model ( N=3 ) for the ferro(antiferro)-magnetism, and the linear sigma model ( N=4 ) relevant for the chiral phase transition in two-flavor QCD. We solve the fRG flow equations at finite temperature across both the symmetric and broken phases, as well as the fixed point equation characterizing the Wilson-Fisher universality class. In all cases, the neural network results show excellent agreement with those obtained from well-established numerical methods, e.g., the Finite Difference Method and the Local Discontinuous Galerkin (LDG). More importantly, it offers a unified framework that naturally handles the stiffness and multi-scale structure of the flow. This paper is organized as follows. In Section II we provide a brief introduction about the fRG, including some theoretical basics, truncation schemes as well as some necessary notations. In Section III we present the fRG flow equations for the O(N) model and the analytical large- N solutions required for our methodology, with detailed derivations collected in App. A . In Section IV we discuss our neural network methodology, including the network architecture, loss

Read the complete text