Can physics-informed neural networks solve functional renormalization group flow equations?

A network that learns the flow itself

The functional renormalization group (FRG) tracks how an effective action changes as fluctuations are integrated out from short to long wavelengths. In the local potential approximation (LPA) used here, that flow reduces to a partial differential equation for the effective potential V_k, which depends on the field invariant rho, the temperature T and the renormalization scale k. Solving it numerically is the hard part: the potential flattens toward convexity in the infrared, and grid-based schemes must discretize the field domain.

The paper's answer is to represent the solution with a neural network and train it by making the flow equation itself the loss. For a candidate function v(rho, T, k) with tunable parameters, the authors compute the derivatives analytically from the network rather than with finite differences, evaluate how badly the candidate violates the flow equation at sampled points, and minimize that residual. This is the physics-informed neural network (PINN) idea: the network is not fitted to data, it is fitted to the differential equation.

Why the large-N shortcut was dropped

An earlier PINN treatment of the same model, by Tan et al., built its baseline from the implicit analytic solution of the effective potential in the limit N to infinity, then trained a network to learn finite-N corrections. That decomposition is clever because the large-N solution already captures most of the infrared flattening, leaving the network only a residual to learn.

The authors argue that this advantage costs generality. A large-N expansion is not necessarily available for more complex multicomponent scalar models, and even if a similar scheme could be constructed, the resulting flow equations are not guaranteed to have an analytic solution. The earlier approach also used finite-difference approximations for derivatives. The present work therefore removes both dependencies: no large-N baseline and no partial reliance on grid techniques.

An ansatz spanning small and large fields

The central construction is a series of ansätze that combine small-field and large-field expansion schemes with a neural network, so that the approximation stays accurate across the entire field range rather than only near rho = 0 or at large rho. The authors describe how they arrived at the final form progressively. In an intermediate version, results were promising except for the flow of the root of V'_k, the minimum point of the potential, when compared with the grid solution; appending rho/rho_0,Lambda to the network inputs resolved that discrepancy.

The network itself is a feed-forward architecture with tanh activations, and the final layer omits the activation so the output is not confined to the interval (-1, 1). Depth and hidden size control how many tunable parameters the ansatz has, and therefore the best accuracy it can represent. The authors solve for the derivative V'_k rather than V_k itself, as the earlier work did, and reconstruct V_k by numerically integrating with respect to rho.

The unstable region forces an infrared cutoff

Some parameter settings make the inverse propagators non-positive at certain (rho, T, k) points, which the authors call the unstable region. Where that happens, the flow equation's right-hand side is undefined because non-positive quantities appear under a square root in the denominator, so the loss cannot be evaluated and training fails. Masking those samples keeps training running, but the residual diverges at the boundary of the unstable region, so the optimizer prioritizes shrinking it.

In the authors' experience, eliminating the unstable region completely comes at the price of accuracy at higher renormalization scales. They therefore introduce an infrared cutoff epsilon > 0 and generate all training triplets with epsilon <= k <= 1, leaving room for residual unstable regions below epsilon. The reported instability threshold, the largest scale the unstable region reaches, sits slightly below epsilon = 0.2, consistent with the optimizer pushing it out of the trained domain. The authors note that maintaining accuracy at k > epsilon while lowering epsilon can be done by increasing model parameters, but that a detailed study is beyond the paper's scope.

At zero temperature the curves nearly coincide

The first test fixes T = 0, the case where flattening of the potential is most prominent, on the reasoning that reproducing the grid result there would give reason to expect the same at higher fixed temperatures. Comparing the grid and PINN solutions over epsilon <= k <= 1, the authors report nearly identical curves in almost all aspects, with the exception of the flow of the quartic coupling g_k.

The comparison also includes a heatmap of the PDE residue in the rho-k plane, with the white region marking where the flow equation's right-hand side is undefined. The instability threshold shown corresponds to the terminal model parameters, not a training hyperparameter: the unstable region and its threshold depend on the model parameters and change throughout training.

Scaling up to learn the full temperature dependence

After the fixed-temperature demonstration with a relatively small network of depth 4 and hidden size 40, the authors extended training to finite and varying temperatures, so the network had to learn V'_k(rho; T) across rho, k and T. They then examined how hidden size affects accuracy. Increasing the hidden size to 100 produced a temperature dependence of the order parameter that was practically indistinguishable from the grid-based result.

This is the paper's strongest claim about scope: the network learns not just a single flow but the joint dependence of the effective potential on field, temperature and scale within the LPA. The comparison standard throughout is the traditional grid-based Runge-Kutta solution, so the evidence is agreement with that reference inside the trained scale window, not an independent analytic benchmark.

Where a discretization-free solver matters

The authors' case for the method rests on what grid methods struggle with. For theories with more complex symmetry groups and multiple invariants, they argue that grid-based methods can become impractical because of inefficient use of available degrees of freedom and numerical instabilities introduced by finite-difference approximations. In the PINN approach there is no discretization, and the network's degrees of freedom are fully decoupled from the domain of the effective potential.

The paper is explicit that this is a suggestion for future work rather than a demonstrated result: extensions to effective potentials depending on multiple field variables, to O(N) x O(M) or SU(N) x SU(N) symmetries, and to gauge theories are listed as natural next tests. The demonstrated evidence remains the O(N) LPA flow at zero and finite temperature, validated against the grid solution.

The cutoff and the reference define the usable range

For a reader deciding whether this method is usable, the decisive variables are the infrared cutoff epsilon and the accuracy target above it. The network is trained only on epsilon <= k <= 1, and the unstable region is deliberately allowed to survive below epsilon; the reported threshold near epsilon = 0.2 is a consequence of that design, not an independent physical prediction. Accuracy at k > epsilon can in principle be improved by increasing width, depth or polynomial order, but the paper does not quantify that trade-off.

The second decisive point is the comparison baseline. Every accuracy claim here is agreement with the grid-based Runge-Kutta solution in the LPA, including the near-identical V'_k curves at T = 0 apart from the quartic coupling flow, and the order-parameter temperature dependence at hidden size 100. The paper's contribution is therefore best read as removing the large-N crutch and the finite-difference derivatives from PINN-based FRG solving while retaining grid-level agreement in the tested O(N) setting, with generalization to multi-invariant theories left as the open test.

Sources used in this answer