Can a language model build a neural network from physical laws?

From equations to networks

Partial differential equations describe how quantities change across space and time. The heat equation, for example, describes diffusion, while the advection equation describes transport. Solving such equations requires finding a field that obeys both the equation and specified starting and boundary conditions.

Physics-informed neural networks, or PINNs, represent that field with a neural network. Training adjusts its numerical parameters, called weights, to reduce violations of the equation and conditions. Those adjustments normally use gradients: calculations indicating how changing each weight affects the objective.

Shogo Sato and colleagues at NTT propose a different construction process, Physics2Weight. A pretrained language model supplies network weights directly and can redesign the network around mathematical structures suggested by the problem. The target network is constructed without gradient-based weight optimization; this does not mean the language model itself was never trained.

Three ingredients

The framework combines explicit physical constraints, architecture editing and repeated weight revision.

First, prompts specify additional requirements, such as bounded amplitudes, symmetry or non-increasing energy. These are checked against stated tolerances rather than assumed to hold because the model received instructions.

Second, the language model can combine small neural-network blocks with mathematical operations. For heat diffusion, it can build exponential decay into the computation. For advection, it can use a coordinate that moves with the traveling wave. Architecture rules limit complexity, require neural-network components and prohibit non-differentiable operations such as conditionals.

Third, an external evaluator checks each candidate and returns equation residuals, initial and boundary errors, coarse-grid predictions and applicable constraint checks. A residual measures how far a proposed solution departs from satisfying the equation when substituted into it.

The language model has no tool access and does not execute these evaluations itself. Five revision rounds produce six candidate weight sets. Final selection prioritizes fewer violated constraints, then a lower physics-informed loss. Reference solutions and errors against them are withheld from the model.

What the experiments found

The researchers tested four equations with one spatial dimension: heat, advection, Burgers’ and Allen–Cahn. The latter two include nonlinear behavior, with a steep viscous shock in the Burgers’ problem and evolving interfaces in Allen–Cahn.

All language-model methods used Claude Opus 5.5. The initial network and PINN baseline each contained 501 parameters. Each experiment had three independent runs; tables report metric medians. Independent solution evaluation used a 256-by-101 space–time grid, with analytical references for heat and advection and numerically computed references for the nonlinear cases.

For heat and advection, Physics2Weight reported relative L2 errors of approximately 4.31 × 10⁻¹⁷ and 2.00 × 10⁻¹⁷, respectively—near floating-point precision. Relative L2 error summarizes the overall discrepancy relative to the reference field’s size. These exceptionally small errors relied heavily on embedding known analytical structures, not simply generating better weights for an unchanged network.

For Burgers’ and Allen–Cahn, relative L2 errors were about 1.53% and 3.53%, compared with 4.69% and 52.0% for the tested PINN. These are aggregate field errors, not guarantees about every location. The generated networks also used fewer parameters than the initial architecture.

Where accuracy and physics diverged

The nonlinear results expose the main caveat. Physics2Weight had higher mean squared equation residuals than the PINN for both Burgers’ and Allen–Cahn, despite lower solution errors.

For Burgers’, abrupt switches within the generated computation produced localized spikes in second spatial derivatives. In one run, increasing spatial sampling from 256 to 2,001 points raised the residual mean square from 0.0545 to roughly 6.6 billion. The PINN’s changed by less than 20%. The authors caution that Physics2Weight’s residual estimates were not grid-converged.

Passing the five specified constraint checks therefore did not establish complete physical consistency. Those checks capture selected properties within tolerances; they do not replace checking the governing equation throughout the domain.

Component comparisons reinforce that distinction. Architecture editing delivered the largest accuracy improvements. Explicit constraint prompting improved constraint satisfaction but did not improve solution error across the four problems. Repeated revision helped some cases, not all, and the full framework was not the most accurate variant on every equation.

What remains uncertain

This is an arXiv preprint, dated November 7, 2026; the supplied text does not establish peer-reviewed publication. Its evidence covers four one-dimensional problems, one language model and three runs per method—not general superiority over numerical solvers or PINNs broadly.

Runtime also varied: fewer network parameters did not guarantee faster construction. The results support investigating physics-guided network design, while showing why solution accuracy, constraint checks and local equation satisfaction must be evaluated separately.