⚛️ phenomenology

Nonperturbative functional renormalization group for Higgs-singlet models with physics-informed neural networks

This paper presents a nonperturbative functional renormalization group framework for Higgs-singlet models that utilizes physics-informed neural networks to solve the Wetterich flow equation without polynomial expansion, successfully reconstructing finite-temperature effective potentials while highlighting the critical necessity and current limitations of soft consistency constraints for selecting physically sensible solutions.

Original authors: Norimi Yokozaki

Published 2026-09-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Norimi Yokozaki

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The universe as we know it is built on a foundation of particles and forces, a structure described by a theory called the Standard Model. For decades, this theory has been remarkably successful, predicting the behavior of matter with stunning precision and accounting for the discovery of the Higgs boson. Yet, despite its triumphs, the Standard Model is incomplete. It cannot explain why the universe is filled with matter rather than an equal mix of matter and antimatter, a mystery known as the baryon asymmetry. To solve this, physicists often look to the very early moments of the universe, specifically a period when the forces of nature underwent a dramatic shift. If this shift happened abruptly rather than smoothly, it could have created the conditions necessary to generate the excess of matter we see today. However, determining whether such an abrupt shift occurred requires calculating the energy landscape of the universe at incredibly high temperatures, a task that pushes the limits of traditional mathematics.

In a recent study, a researcher at Zhejiang University has tackled this difficult calculation using a novel approach that blends advanced physics with artificial intelligence. The goal was to map out the "effective potential," a concept that describes the energy state of the universe's fields as they cooled down after the Big Bang. Specifically, the study focused on a version of the Standard Model that includes an extra, invisible particle called a singlet. This extra particle is a popular candidate for explaining dark matter and could potentially trigger the abrupt shift needed to create the matter-antimatter imbalance. The challenge is that the equations governing this shift are incredibly complex, involving multiple variables that change simultaneously as the universe cools. Traditional methods for solving these equations often rely on simplifying assumptions or breaking the problem into small, rigid grids, which can miss subtle but crucial details or fail to converge on a solution entirely.

To overcome these hurdles, the researcher developed a new framework that uses a type of artificial intelligence known as a physics-informed neural network. Instead of forcing the solution onto a fixed grid, this method treats the energy landscape as a continuous surface, allowing the computer to learn the shape of the solution directly from the underlying physical laws. The neural network acts as a flexible, mesh-free map that can adapt to the changing conditions of the early universe without the distortions that come from rigid mathematical grids. By training this network to satisfy the fundamental equations of the functional renormalization group—a powerful tool that tracks how physical laws change with scale—the researcher was able to generate a smooth, continuous description of the energy potential across the entire range of temperatures and field values.

The study applied this method to a specific scenario where the universe transitions from a symmetric state to a broken state, a process relevant to the electroweak phase transition. The researcher tested the system at two different temperatures, 100 GeV and 200 GeV, and also reconstructed a two-dimensional view of the energy landscape at 100 GeV. The results showed that the neural network could successfully navigate the complex flow of the equations, providing a detailed picture of how the potential evolves. However, the study also revealed a significant limitation: the equations are so sensitive that the network sometimes drifted toward mathematically valid but physically nonsensical solutions. To prevent this, the researcher had to introduce a "soft constraint," a guiding rule that kept the network's output close to what is expected from simpler, well-understood theories. Even with this guidance, the final result retained a small dependence on how this rule was tuned, indicating that while the method works, finding a completely independent way to select the correct physical solution remains an open challenge.

The researchers compared their new neural network approach against two other methods: a traditional grid-based solver and standard perturbation theory, which is a common approximation technique. The grid-based solver, which attempts to solve the equations by stepping through a fixed lattice of points, struggled significantly. In many cases, it failed to find a stable solution, or it converged to a result that looked mathematically correct but was physically impossible, such as a potential that dropped to unrealistic depths. The neural network, by contrast, provided a continuous and stable description, avoiding the jagged artifacts and convergence failures of the grid method. Yet, the study was careful to note that the neural network did not simply replace the old methods; it required the same physical guidance to ensure it stayed on the right track. The comparison showed that while the neural network could reproduce the general behavior of the system, the precise details still relied on the external guidance provided by the soft constraint.

This work represents a proof of concept for using artificial intelligence to solve some of the most difficult equations in theoretical physics. It demonstrates that neural networks can handle the multi-dimensional, non-linear problems that arise when studying the early universe, offering a more flexible alternative to traditional grid-based calculations. The study successfully mapped out the energy landscape for a model with an extra singlet particle, showing how the potential changes as the universe cools. However, it also highlighted that the method is not yet a fully autonomous solver. The need for the soft constraint suggests that the equations themselves contain ambiguities that current mathematical techniques cannot resolve on their own. The researcher identified this reliance on external guidance as the central problem to be solved in future work.

Ultimately, the study provides a clearer, more continuous view of the conditions required for a strong first-order phase transition in the early universe. Such a transition is a necessary ingredient for theories that explain the origin of matter. By showing that a neural network can successfully model this transition, the research opens a new path for exploring the physics of the early universe. The findings suggest that while the tools are becoming more powerful, the complexity of the physical laws governing the cosmos still demands a careful balance between computational innovation and established physical principles. The work does not claim to have solved the mystery of the matter-antimatter asymmetry, but it has provided a more robust way to test the models that might explain it, paving the way for future investigations into the hidden sectors of particle physics.

Technical Summary: Nonperturbative Functional Renormalization Group for Higgs–Singlet Models with Physics-Informed Neural Networks

Problem Statement
The Standard Model (SM) fails to explain the observed baryon asymmetry of the universe (BAU), necessitating extensions such as the addition of a real $Z_2$ -symmetric singlet scalar. A viable mechanism for baryogenesis in these extensions requires a strongly first-order electroweak phase transition (EWPT). However, calculating the finite-temperature effective potential for such multi-field systems is technically challenging. Conventional perturbative methods suffer from severe infrared (IR) sensitivity and resummation issues near the critical temperature. While the Functional Renormalization Group (FRG), formulated via the Wetterich equation, offers a nonperturbative solution, solving the resulting nonlinear partial differential equations (PDEs) in multi-field space is difficult. Standard approaches relying on low-order polynomial truncations or grid-based discretizations struggle with the continuous, multi-dimensional nature of the flow, often failing to converge or requiring ad-hoc approximations.

Methodology
This work develops a nonperturbative FRG framework within the Local Potential Approximation prime (LPA') to solve the Wetterich flow equation for the $Z_2$ -symmetric real singlet extension of the SM at zero and finite temperature. The core innovation is the use of Physics-Informed Neural Networks (PINNs) to represent the scale-dependent effective potential without polynomial expansion or fixed grids.

Framework: The author employs a hybrid ansatz where the effective potential $u(\rho, \sigma)$ is decomposed into a physics-motivated baseline (running tree-level potential and a crude thermal mass term) and a learnable residual represented by a neural network $N(t, \rho, \sigma)$ .
Network Architecture: The network is designed to output the potential and its first and second derivatives ( $N, N_\rho, N_\sigma, N_{\rho\rho}, N_{\sigma\sigma}, N_{\rho\sigma}$ ) simultaneously via a modular architecture with shared and branch-specific layers. This avoids repeated automatic differentiation, enhancing efficiency and stability.
Flow Equation: The flow is governed by the dimensionless Wetterich equation, incorporating contributions from Goldstone modes, Higgs-singlet radial modes, and external perturbative sectors (top quark and electroweak gauge bosons) via a term $E_{ext}$ . Anomalous dimensions ( $\eta_H, \eta_S$ ) are included at the LPA' level using perturbative one- and two-loop results.
Training Strategy: The solution is obtained by minimizing a loss function that includes the PDE residual (enforced via automatic differentiation), boundary conditions, and a "soft consistency constraint." This constraint penalizes deviations from perturbative expectations in sign and order of magnitude near the field-space origin, acting as a guide to select physically sensible solutions.
Benchmarking: The PINN approach is validated against two baselines: (1) resummed perturbation theory (including ring resummation) and (2) a grid-based Newton–Krylov relaxation solver. The grid solver was found to struggle with convergence, particularly in the presence of thermal loops and near symmetry axes, often converging to unphysical solutions or stalling.

Key Results

Continuous Description: The PINN framework successfully provides a continuous, mesh-free description of the effective potential over the full relevant field and scale range, avoiding the artifacts of grid interpolation.
Phase Transition Analysis: The framework was applied to benchmark temperatures ( $T=100$ GeV and $T=200$ GeV). It successfully reconstructed one-dimensional field-space slices; however, the two-dimensional reconstruction at $T=100$ GeV is identified as the "least converged output" of the network, with residual systematic errors regarding the inner barrier region.
Bounce Action: While the framework allows for the evaluation of the potential, the full extraction of the two-step bounce action and the determination of the nucleation temperature ( $T_n$ ) via a temperature scan are explicitly left for future work due to the computational cost of retraining the network at each temperature.
Solver Comparison: The grid-based Newton–Krylov solver failed to converge to the requested tolerance ( $10^{-8}$ ) for the full two-field system, often stalling at residual norms of order $10^{-1}$ or converging to spurious deep wells. In contrast, the PINN approach yielded stable solutions, though they retained a dependence on the soft consistency constraint.
Role of Constraints: The study found that the soft perturbative-consistency constraint is not merely helpful but necessary for selecting a physically sensible solution. Without it, the network could converge to mathematically valid but physically incorrect trajectories. However, the final results showed a residual dependence on the specific weights and analytic targets used for this constraint.

Significance and Claims
The paper presents a proof-of-concept for applying genuinely continuous, autodiff-based PINNs to multi-field, finite-temperature FRG flow equations, a task previously unexplored for phenomenologically motivated Beyond-the-Standard-Model (BSM) extensions.

Methodological Advance: It demonstrates that PINNs can handle the nonlinearity and multi-dimensional complexity of the Wetterich equation more effectively than traditional grid-based relaxation methods, which suffer from conditioning issues and convergence failures in the presence of thermal loops.
Physical Insight: The work highlights that the nonlinear residual of the flow equation possesses spurious near-stationary points. Consequently, a small numerical residual does not guarantee a physically correct solution, necessitating the use of external guidance (the soft constraint) to anchor the solution to physical expectations.
Limitations and Open Problems: The author modestly notes that the converged results retain a residual dependence on the hand-tuned weights and the analytic thermal target used in the consistency constraint. They identify the elimination of this dependence—achieving a solution that is robust without such guidance—as the central open problem for future work.
Future Directions: The paper outlines that while the current model (real singlet) establishes the methodology for a strongly first-order transition, extending the framework to include CP-violating sectors (complex singlets or additional interactions) is required to fully address the baryogenesis mechanism.

In summary, this work establishes a viable, nonperturbative computational pipeline for studying electroweak phase transitions in scalar-extended SMs, overcoming specific numerical hurdles of traditional methods while transparently identifying the current reliance on perturbative priors as a limitation to be addressed.

Technical Summary: Nonperturbative Functional Renormalization Group for Higgs–Singlet Models with Physics-Informed Neural Networks

More like this