Neural networks vs finite elements: 10,000 points—not a drift test

The supplied file gives 49.87% without defining what the percentage measures. It reports no shared test distribution, matched boundary conditions, mesh-refinement results, finite-element tolerances, or common accuracy metric. A clean percentage can look decisive while remaining impossible to audit. The evidence cannot establish whether neural networks or finite elements better withstand structural shift after deployment.

That gap matters because physics-informed neural networks embed governing equations and continuous coordinate fields, potentially helping when labeled data are scarce. Those constraints can improve generalization, but do not test whether frozen weights retain accuracy as regimes move. The comparisons define no frozen-model boundary, update cutoff, or post-deployment window. Any claimed 49.87% result could depend on favorable data, hidden retraining, or incomparable settings unless its protocol is explicit.

For seismic design, nonlinear finite-element analysis should remain the default while neural surrogates are qualified, not installed as replacements. A candidate must reproduce matched conditions, expose mesh and solver sensitivity, and report a common metric. It should then face frozen-weight structural shifts and an amortized lifecycle speed test, with bias measured against the engineering margin. Point count is not a drift test: neither many sampled points nor embedded physical laws proves a model will preserve decisions when structure, loading, or context changes.

Raissi's Cavity-Flow Benchmark Is Not a 2026 Drift

Raissi’s cavity-flow benchmark is evidence about the method, not a 2026 qualification of a frozen seismic PINN. It shows that a residual-trained network can solve a bounded PDE; it does not show that fixed weights retain accuracy when stiffness, damping, loading, or earthquake-record shapes change. That distinction controls replacement: in-distribution field accuracy is not evidence of no-retraining generalization.

According to Raissi, Perdikaris, and Karniadakis’s JCP report, their 9-hidden-layer, 20-neuron-per-layer PINN reached roughly 10^-3 cavity-flow field error using interior collocation points. I use that result as mechanism-validation evidence: the architecture and residual formulation can recover a canonical flow solution. It is not evidence of structural extrapolation, nonlinear hysteresis, or frozen inference under record shift. Treating interpolation in one PDE as qualification for another would be a category error.

The efficiency evidence points the same direction. According to Grossmann, Vabnik, Schömer, and Elser’s unit-square 2-D Poisson comparison, PINNs showed no consistent wall-time advantage at matched solution error. I treat this as direct solver-efficiency counter-evidence, not proof that every PINN is slow. It does mean an unsupported neural speedup cannot be imported into a structural decision: preprocessing, inference, and postprocessing must together beat nonlinear finite-element analysis at the declared end-to-end timing boundary.

Seismic qualification is a vector problem, not a displacement-field contest. According to Ibarra and Krawinkler’s JSE study, 9-, 12-, and 20-story reinforced-concrete frames were benchmarked for global collapse. The study establishes peak drift, hinge state, and force response as quantities a qualified surrogate must reproduce. A low displacement L2 error can conceal a base-shear error or incorrect hinge state, so field error alone cannot determine seismic acceptability.

According to the ASCE 41-17 concrete-moment-frame acceptance row, the transient-drift limits distinguish Immediate Occupancy, Life Safety, and Collapse Prevention. These are decision boundaries, not training regularizers. A visually smooth displacement field can still cross the wrong boundary; blind evaluation must therefore test limit-state classification rather than contour appearance.

According to the Pacific Earthquake Engineering Research Center, the NGA-West2 database contains a broad set of ground-motion records. That population is a warning against treating a few familiar earthquakes as a robustness test. Record, frequency-content, duration, and amplitude shift must be represented in predeclared blind queries. A random split around an already-trained PINN does not do that: it resamples the same data manifold instead of testing unseen structural or seismic shapes with frozen weights.

The practical 2026 move is to freeze the weights, blind queries, tolerances, hardware, and timing boundary before evaluation, then compare the PINN and nonlinear finite-element reference on identical tasks. Only an all-gates pass earns replacement status.

The 2026 Scorecard

Overall verdict: I name nonlinear finite-element analysis the default for safety-critical design and code-compliance work. According to the supplied-source audit, no supplied comparison defines a frozen-model/no-retraining boundary, update cutoff, or post-deployment test window; no supplied source states that weights remained frozen. The evidence therefore establishes neither method as more drift-resistant without retraining. A random split around a trained PINN cannot repair that omission: with weights fixed, it samples the represented manifold, not unseen stiffness, damping, load, or record shapes. The rows below are veto gates, never an average.

Freeze the weights before querying the predeclared blind suite. Bind query identifiers, inputs, model version, acceptance outputs, and any fallback decision into a test manifest. The accuracy limits below are governance thresholds for this decision, not universal laws of PINNs. A code-limit acceptance-band flip is a discrete failure, not a rounding issue that can be diluted by a smaller mean error.

According to the alphaXiv overview of arXiv:2511.14348, a PINN method can enforce hidden irreversible-process laws as soft constraints during training rather than relying only on explicit PDE-residual minimization. The transferable lesson concerns representation, not automatic qualification: encoding a law does not prove that a structural query’s active mechanism and state are covered by a frozen model.

Lifecycle timing must begin before setup and retain inference, monitoring, failed blind queries, and FE fallback; a fallback cannot be discarded as “exceptional” when it is the designed response to drift. I name the PINN overall only if it passes blind-shift accuracy and runtime without losing mechanics, state coverage, or auditability. Every remaining tie goes to nonlinear FE.

Beyond the Envelope

For current structural and seismic qualification, optimization is an audited result, not a reward for expressive architecture. According to Wang, Teng, and Perdikaris’s 2021 gradient-pathology work and Krishnapriyan and colleagues’ 2021 failure-mode study, PINN optimization can fail even when an exact PDE solution is known. I therefore treat convergence as a separate qualification risk: network capacity does not guarantee a mechanically admissible seismic state.

A small PDE residual plus boundary loss provides no universal bound on displacement, reaction, stress, or energy error. With soft constraints, an optimizer can reduce aggregate loss while a local point violates equilibrium, an interface violates continuity, or a component crosses an engineering limit state. I audit fields and derived quantities directly, rather than infer acceptability from the loss scalar. The loss is an optimization signal, not proof of mechanics after a blind shift.

Nonlinear FE is not an infallible oracle. Mesh density, damping, section cracking, bond slip, contact treatment, and soil idealization can move the reference result. The supplied-source audit reports no shared test distribution, identical initial or boundary conditions, mesh-refinement results, finite-element solver tolerances, or common accuracy metric. I require an independent convergence and calibration audit before declaring FE the winner; otherwise nonlinear FE remains the reference choice, not a defeated default.

Even a clean blind run has a statistical ceiling. The qualification arithmetic below converts a perfect observed run into a bounded lower-confidence statement, not future-proof reliability. It cannot authorize a frozen parametric PINN without the predeclared blind-shift tolerances, engineering-error gates, code-limit checks, and required end-to-end speedup. If any condition fails, nonlinear FE is the choice.

Extrapolation is mechanism-dependent, not average-dependent. A stable small-strain cantilever can coexist with brittle buckling, low-cycle fatigue, fracture, or soil–structure interaction in one nominal model class. I report the worst behavior within each mechanism family rather than only mean error, so a benign family cannot hide a brittle tail. The tempting random in-distribution split around a trained PINN fails here: it samples the same stiffness, damping, load, and record shapes, not unseen behavior with frozen weights. No-retraining claims must survive that shift.

Structural-health-monitoring deployments add observability error. Sensor bias, corrosion, missing channels, and changing excitation can resemble Δξ even when the mechanism has not changed. I separate instrumentation faults from Δξ by checking channel health, excitation consistency, and residual signatures before authorizing retraining or a solver change. This prevents a measurement artifact from being mislabeled as model drift—and prevents a PINN from being frozen or FE being rejected for the wrong reason.

What the Data Doesn't Tell You

The rule does not break when a frozen PINN fails; the evidentiary chain breaks before the rule can be invoked. On the present record, nonlinear finite-element analysis remains the defensible choice. According to the alphaXiv overview published on 11 December 2025, with Sifan Wang shown as author, the available source is an overview rather than a case-resolved blind-shift audit. It can motivate evaluation, but it cannot expose the tail errors, failed classifications, and timing distributions needed to support replacement.

Limitations of the evidence: An ordinary random train/test split around a trained PINN is not certification evidence. It samples the same data manifold and can reward interpolation in stiffness, damping, loading, and record shape; it does not test frozen weights after a genuine blind shift. Residual loss and low in-distribution response error say nothing by themselves about localized drift peaks, code-limit decisions, or end-to-end latency. An overview also removes access to per-query outputs, failed runs, and the timing boundary. Without a case ledger, a favorable aggregate cannot establish that the predeclared blind-suite conditions were met.

Variance across cases: The decisive diagnostic is dispersion, not a single headline error. A structure well away from a governing limit may tolerate a small response perturbation, whereas a nearby structure can flip a code classification after a smaller one. Localized drift peaks and strongly nonlinear response can be diluted by displacement-level summaries. Runtime can also change with numerical tolerances, preprocessing, hardware, and solver behavior; a speed claim obtained on favorable hardware or by excluding required work is not the same end-to-end comparison. I would require a per-case record linking frozen weights, shifted inputs, each gate result, code decision, timing boundary, and execution environment before accepting the suite verdict.

Worked OOD Rejection

The defensible outcome is rejection before response inference. I use a dimensioned validation case based on the beam equations in Timoshenko and Goodier’s Theory of Structures, second edition: a 5.0 m steel cantilever with b = 0.30 m, h = 0.50 m, a baseline modulus E, and ν = 0.30. I present this as analytical benchmark evidence, not as a field experiment or a claimed PINN response.

I lock the fixed-section PINN training envelope to P = 40–80 kN and a bounded range for E. I then submit P = 90 kN and a value of E within that locked range with the network and its scalers frozen. The blind stiffness remains inside the locked range, but the load does not. The out-of-distribution monitor must classify the load before I inspect any predicted response. That ordering prevents a plausible displacement from laundering an inadmissible query. A random holdout within the training envelope cannot substitute for this test: it samples the same data manifold rather than a shifted load.

At the blind point, I calculate the following reference quantities according to the equations in Timoshenko and Goodier. These values characterize the benchmark; they are not substituted for a PINN prediction.

The finite-element comparison is a numerical verification target for the analytical beam response, not evidence that the frozen PINN is admissible. Here, 90 kN is 12.5% beyond the 80 kN envelope maximum. I therefore mark the PINN not qualified before inference and select nonlinear finite-element analysis as the explicit winner. I do not invent a PINN displacement to turn an out-of-envelope query into a success. The actionable control is simple: classify the shifted input first, and preserve that rejection even if the rejected model would be fast.

Choose in Five Rules

For the current qualification, a random in-distribution holdout around a trained PINN is not a deployment test: it samples the same data manifold and leaves unseen stiffness, damping, load, and record shapes unresolved. I therefore use a fail-closed decision tree in which nonlinear finite-element analysis (FE) is the default. “PINN eligible” means the complete frozen, no-retraining pipeline—not a network repaired after the query is known.

I evaluate the rows in order and stop at the first FE trigger. The cited long-term prediction work clarifies the boundary: according to the 2023 Journal of Computational Physics paper on weight-adaptive physics-constrained AI, volume 474, article 111722, DOI 10.1016/j.jcp.2022.111722, adaptive weighting addresses conventional PINNs’ dependence on manually specified loss weights. That is methodological context, not permission to tune a deployed model for each query.

What to do next