Lattice fermion formulation via Physics-Informed Neural Networks: Ginsparg-Wilson relation and Overlap fermions

This paper proposes a Physics-Informed Neural Network framework that formulates lattice fermions as an optimization problem, successfully reconstructing the overlap fermion operator and autonomously discovering both standard and generalized Ginsparg-Wilson relations without relying on predefined approximations.

The Gist

Imagine you are trying to build a perfect digital map of a city, but there's a catch: the rules of physics say you can't draw the city without accidentally creating "ghost" versions of every building. In the world of particle physics, these "ghosts" are called fermion doublers. For decades, physicists have struggled to create a mathematical map (called a lattice) of subatomic particles that is accurate, doesn't create these ghosts, and still respects the delicate rules of symmetry.

This paper introduces a new tool to solve this puzzle: Physics-Informed Neural Networks (PINNs). Think of this not as a human trying to solve a complex equation with a pencil, but as a highly disciplined AI student that learns the rules of the universe by trial and error.

Here is a breakdown of what the authors did, using simple analogies:

1. The Problem: The "Ghost" City

In the past, physicists had to manually design the rules for these particles. They faced a "No-Go" theorem, which is like a sign saying, "You cannot have a map that is local (close-range), symmetrical, and ghost-free all at once."

• The Old Way: Physicists had to choose which rule to break. They would sacrifice symmetry to get rid of ghosts, or sacrifice locality to keep symmetry. It was a game of "pick your poison."
• The New Way: The authors propose letting a Neural Network (an AI) figure out the best compromise. They don't tell the AI the answer; they just give it the "laws of the land" (physical constraints) and let it find the path.

2. The Method: The "Soft Constraint" Coach

The authors trained the AI using a system of "soft constraints." Imagine a coach training an athlete. Instead of saying, "You must run exactly this fast," the coach says, "If you run too slow, you get a small penalty. If you run too fast, you get a penalty. If you trip, you get a big penalty."

• The Penalties (Loss Functions):
  • Symmetry Penalty: If the AI breaks the rules of chiral symmetry (a specific type of particle balance), it gets a penalty.
  • Locality Penalty: If the AI's map connects points that are too far apart (like a teleportation spell), it gets a penalty. The goal is to keep connections local, like neighbors talking to neighbors.
  • Ghost Penalty: If the AI accidentally creates "ghost" particles (doublers), it gets a heavy penalty.

3. Achievement #1: Learning the "Overlap" Map

First, the authors gave the AI a specific target: the Ginsparg-Wilson (GW) relation. This is a famous, complex mathematical rule that allows particles to exist without ghosts while keeping symmetry.

• The Result: The AI successfully learned to recreate the Overlap Fermion operator.
• The Analogy: Usually, to calculate this operator, humans have to use a complicated "recipe" involving long lists of numbers (polynomials or rational approximations). The AI didn't need the recipe. It learned the "shape" of the solution directly. It figured out how to turn a "rough" map (the Wilson kernel) into a "perfect" map (the Overlap operator) just by trying to minimize its penalties. It did this with high precision in both 2D and 4D (simulating our 3D space plus time).

4. Achievement #2: The AI Discovers the Rules Itself

This is the most surprising part. Usually, you tell the AI, "Here is the GW rule, please follow it." But in this second experiment, the authors said, "We don't know the rule yet. Here is a blank canvas of possible mathematical shapes (a generalized polynomial). You figure out which shape works."

• The Setup: The AI was allowed to mix and match different mathematical terms (like mixing ingredients in a soup) to see what happened.
• The Discovery:
  1. Standard Solution: When the AI started with no bias, it naturally figured out that the simplest, most effective solution was the standard GW relation. It essentially "discovered" the famous rule on its own, suppressing all the complicated extra terms that weren't needed.
  2. The "Fujikawa" Solution: When the researchers slightly nudged the AI's starting point (like giving it a tiny hint to look at a different ingredient), the AI found a different valid solution. This solution corresponded to a "Generalized GW relation" proposed by a physicist named Fujikawa.

5. Why This Matters (According to the Paper)

The paper claims this is a shift from "human analytical ingenuity" to "machine-assisted algebraic discovery."

• The Metaphor: For decades, humans have been the only ones capable of solving these complex algebraic puzzles. This paper shows that an AI can not only solve the puzzle but can also explore the "landscape" of possible solutions to find different, valid mathematical structures that humans might have missed or found too difficult to derive manually.

Summary

The authors built a digital "playground" where a Neural Network was tasked with building a particle physics model.

1. They showed the AI could learn a known, perfect solution (Overlap fermions) just by being told to avoid ghosts and stay local.
2. More importantly, they showed the AI could invent the mathematical rules itself. Starting from scratch, it derived the standard rules of the game, and with a slight nudge, it found a new, valid variation of the rules (the Fujikawa relation).

The paper concludes that this method opens a new door for discovering fundamental mathematical structures in physics, potentially finding new ways to describe the universe that we haven't thought of yet.

Technical Summary

Technical Summary: Lattice Fermion Formulation via Physics-Informed Neural Networks

Problem Statement
Formulating fermions on a discretized spacetime lattice is a fundamental challenge in Quantum Chromodynamics (QCD), constrained by the Nielsen-Ninomiya no-go theorem. This theorem dictates that a single-flavor lattice Dirac operator cannot simultaneously satisfy locality, translational invariance, hermiticity, and exact chiral symmetry. Historically, solutions have required sacrificing specific conditions (e.g., Wilson fermions break chiral symmetry; staggered fermions complicate flavor interpretation). The Ginsparg-Wilson (GW) relation offers a paradigm shift by deforming the chiral symmetry algebra, leading to exact solutions like the overlap fermion. However, these exact solutions present significant practical hurdles: they require the evaluation of a matrix sign function, typically approximated via computationally expensive Chebyshev polynomials or Zolotarev rational approximations, and often suffer from delicate locality issues dependent on background gauge fields.

Methodology
The authors propose a novel framework utilizing Physics-Informed Neural Networks (PINNs) to construct lattice fermion operators and discover their underlying algebraic structures. Instead of analytical derivation, the formulation of the Dirac operator is treated as a constraint-driven optimization problem.

1. Neural Network Architecture: The Dirac operator $D(k)$ is parameterized by a Multi-Layer Perceptron (MLP) defined in momentum space. The network takes lattice momentum functions (e.g., $\sin k_\mu$, Wilson kernel terms) as input and outputs the components of the Dirac operator. Parity symmetry is enforced by symmetrizing the network's forward pass.
2. Soft Constraints (Loss Function): The training process minimizes a total loss function composed of weighted "soft constraints":
   • Symmetry/Algebraic Constraints ($L_{GW}$ or $L_{pol}$): Enforcing the GW relation ($D\gamma_5 + \gamma_5 D = D\gamma_5 D$) or a generalized polynomial ansatz.
   • Locality Constraints ($L_{loc}$): Penalizing long-range real-space hoppings via Fast Fourier Transforms (FFT) to encourage exponential decay $|D(r)| \sim e^{-\alpha|r|}$.
   • Spectrum/Pinning Constraints ($L_{pin}$): Ensuring the physical mode remains massless while pushing unphysical doublers to the cutoff scale.
3. Two-Phase Approach:
   • Phase I (Operator Construction): The GW relation is provided as a target constraint. The network learns to map a Wilson kernel to an overlap-like operator.
   • Phase II (Algebraic Discovery): The explicit GW constraint is removed. Instead, a generalized polynomial ansatz $D + D^\dagger = \sum c_n (D^\dagger D)^n$ is used. The network autonomously optimizes the coefficients $c_n$ (using a Differentiable Architecture Search-inspired mechanism with Softmax gating) to satisfy locality and doubler-decoupling conditions.

Key Contributions and Results

• Reproduction of Overlap Fermions: When trained with the GW relation as a soft constraint, the neural network successfully reproduces the overlap Dirac operator to high numerical accuracy in both 2D and 4D. Crucially, the network learns an effective sign-function mapping that forms a sharp step-like profile without explicitly using prescribed polynomial or rational approximations. The learned operator exhibits the expected exponential spatial locality.
• Autonomous Discovery of the Standard GW Relation: In the absence of a pre-defined GW constraint, the network, starting from an agnostic generalized polynomial ansatz, autonomously drives higher-order terms to zero and converges to the standard GW relation ($c_1 \approx 1, c_{n>1} \approx 0$). This demonstrates that the standard GW structure emerges as the preferred solution within the optimization landscape when guided by locality and doubler-decoupling requirements.
• Discovery of Generalized GW Relations: By introducing a slight initial bias in the search space (specifically toward the $n=2$ term), the framework discovers a distinct, stable solution corresponding to a Fujikawa-type generalized GW relation ($D + D^\dagger = \frac{1}{4}(D^\dagger D)^2$). This indicates that the optimization landscape contains multiple physically valid algebraic structures depending on the initial search bias.

Significance and Claims
The paper claims to establish a new machine-learning route for the fundamental formulation of lattice field theories. The authors assert that their PINN framework serves as a complementary tool to traditional analytical methods and previous optimization programs (such as perfect actions).

The primary significance lies in the transition from operator construction to machine-assisted algebraic discovery. The study demonstrates that deep learning models can not only approximate known solutions (like the overlap operator) but can also autonomously derive the underlying algebraic constraints (the GW relation and its generalizations) solely from physical requirements such as locality and the suppression of fermion doublers. The authors suggest this approach could be extended to Hamiltonian formalisms, minimally doubled fermions, and systems with background gauge fields, potentially revealing new classes of localized operators and optimizing kernels for specific symmetries.

The work remains modest in scope, noting that current 4D results utilize small lattices ($L=10$) and that future work is required to investigate locality properties and bare kernels more thoroughly in higher dimensions. The framework is presented as a proof of concept for using differentiable programming to explore the space of valid lattice actions.