Neural Surrogates for Relativistic Systems

A neural surrogate can evaluate a relativistic waveform or field map far faster than the calculation that generated its training data. Scientific usability depends on more than median test loss: the model must preserve the symmetries and constraints relevant to its output, expose its training and inference domain, remain stable across seeds and implementation choices, and fail visibly at boundaries and under distribution shift. Speed is evidence only at a matched physical accuracy target.

Required background. Probability and Inference for Gravity supplies train/validation/test separation, calibration, and uncertainty metrics.

Helpful background. Formulations of General Relativity supplies constraints, gauge content, and physical symmetries; Differential Equations for Relativity supplies high-fidelity residual and convergence checks; Computational Relativity supplies reproducible benchmarks and failure diagnostics.

What the network is allowed to learn

Let $h(t;\lambda)$ be a versioned high-fidelity waveform at parameters $\lambda$. A practical surrogate often factors the problem as

$$h(t;\lambda) \simeq \sum_{j=1}^{n}c_j(\lambda)e_j(t), \qquad c(\lambda)\simeq F_\theta(\widetilde\lambda),$$

where $e_j$ is a frozen reduced basis, $\widetilde\lambda$ is a declared scaled coordinate, and $F_\theta$ is a neural map. This division keeps the output dimension small and allows basis-projection error to be measured separately from regression error. A direct field network needs the same separation through an independently evaluated continuum residual and constraint norm.

Preprocessing defines part of the model. Time and phase alignment, amplitude normalization, parameter transformations, mode selection, tapering, and whitening can remove variation or introduce discontinuities. Fit these operations on training data only and version them. A model trained on $h_{22}$ under $h=h_+-ih_\times$ cannot silently be applied to a different polarization sign, harmonic normalization, or time convention.

Relevant structure depends on the target:

• exchange and label symmetries in binary parameters;
• spin-weight and rotation behavior of waveform modes;
• equatorial-reflection relations such as $h_{\ell,-m}=(-1)^\ell h_{\ell m}^*$ only when the source and preprocessing preserve that symmetry;
• positivity, causality, conservation, or hyperbolicity for closures and evolution maps;
• Hamiltonian, momentum, gauge, or divergence constraints for fields;
• known weak-field, test-mass, equal-mass, or symmetry limits.

A structure can be enforced by representation, architecture, loss, or post-processing. Each choice needs a separate residual test because a soft penalty can be violated and a projection can hide the size of the raw error.

Khan and Green describe a neural waveform surrogate built from reduced data in Khan and Green 2021, § II, Eqs. (1)–(10), and §§ III.A–III.C. The architecture is one design point, not a universal prescription.

What coefficient error means for waveform error

If the frozen basis is orthonormal in the same norm used to score the waveform, let $c_j=\langle e_j,h\rangle$ and let $\widehat c_j$ be the network prediction. Orthogonality gives an exact decomposition:

$$\|h-\widehat h\|^2 =\|h-P_nh\|^2+\sum_{j=1}^n|c_j-\widehat c_j|^2.$$

The cross term vanishes because $h-P_nh$ is orthogonal to every basis vector. Thus a projection floor of $10^{-3}$ and a coefficient error of $2\times10^{-3}$, measured in units where $\|h\|=1$, produce total error $\sqrt5\times10^{-3}$. More network training cannot remove the projection contribution.

This identity fails in that form if coefficient scaling, basis nonorthogonality, or waveform weighting changes; then include the appropriate Gram matrix. It also does not equate coefficient error with detector mismatch. For example, $\widehat h=1.1h$ has a 10% relative amplitude error but zero normalized mismatch. Report an amplitude-sensitive norm alongside normalized overlap.

Comparison of methods and results

The evidence below was checked through 2026-08-18.

Position	Evidence in its tested regime	Boundary of the claim
established baseline	reduced-basis and empirical-interpolation surrogates provide explicit projection and node errors	their parameter fits and speed still require held-out validation
useful neural result	Khan and Green used $2\times10^5$ training examples and $2\times10^4$ examples in each of the validation and test sets; they reported median mismatch about $2\times10^{-5}$	the test family, preprocessing, modes, parameter box, detector norm, and architecture delimit the result
adverse evidence within the same study	worst mismatches reached about $2\times10^{-3}$ and clustered near parameter boundaries	median accuracy does not control boundary risk or a downstream worst-case requirement
performance evidence	fast inference was demonstrated for the stated hardware and implementation	preprocessing, transfer, warm-up, batching, and comparison baseline must accompany any speedup
not established	reliable extrapolation, universal uncertainty, preservation of untested constraints, or superiority to classical surrogates at equal worst-case error	each needs independent domain-specific evidence

The data design and training choices are documented in Khan and Green 2021, §§ III–IV and Tables I–II; the mismatch distribution, boundary failures, and timing appear in Khan and Green 2021, § V, Figs. 2–3 and Table III. The favorable position is that a neural map can give a useful controlled speed gain inside a demonstrated domain. The competing position is that seed sensitivity, boundary concentration, missing constraints, and distribution shift prevent transferring that result to a broader physical claim.

A benchmark that makes worst cases visible

Freeze a compact-binary waveform family, source revisions, modes, parameter box, time or frequency interval, alignment, norm, and reduced basis. Split by physical parameter point before preprocessing into scattered interpolation, complete boundary faces and corners, a contiguous withheld interior region, and out-of-domain tests. Never split time samples from one waveform across training and test sets.

Train at least five declared seeds and compare:

1. the neural coefficient map;
2. a classical empirical-interpolation or polynomial/regression surrogate; and
3. a Gaussian-process or distance-aware baseline when uncertainty is claimed.

For every method, report median, 95th-percentile, and maximum detector-weighted mismatch; maximum phase and relative-amplitude error over the declared interval; basis projection and regression contributions; exact symmetry residuals; and the worst parameter location. For a field or closure, add the original continuum equation, conservation, positivity, and constraint residuals evaluated by a path independent of the training loss.

Increase training coverage, network width or depth, optimization accuracy, reduced-basis size, and arithmetic precision separately. Report variation across seeds, not only the best checkpoint. A representative acceptance contract might require maximum in-domain mismatch below $10^{-3}$, 95th percentile below $10^{-4}$, exact discrete symmetries at roundoff when enforced by construction, and no seed whose boundary error exceeds the threshold. Actual discrepancies must be preserved.

The adverse check removes a contiguous parameter region touching a physical boundary, trains without it, and measures error while approaching and crossing that boundary. The model must either remain within the declared tolerance inside the validated domain or signal refusal; confident smooth continuation is not success when the high-fidelity error is large. The independent comparison reconstructs held-out waveforms directly and recomputes mismatch using separately implemented quadrature.

Classical surrogate construction offers a useful control because it exposes basis, interpolation, fitting, and online errors separately; the offline–online decomposition and held-out error tests are given in Field et al. 2014, §§ III–V and Figs. 9–11. A neural model should be preferred only after matching domain, data, waveform representation, error target, and hardware.

Physical structure and uncertainty

Adding a constraint penalty

$$L(\theta)=L_{\rm data}+\alpha L_{\rm constraint}$$

does not guarantee a constrained output. The scale $\alpha$, sampling measure, derivative accuracy, and trade with data fit determine the result. Evaluate the constraint on a dense independent set, including domain boundaries, and compare with an architecture that enforces the relation exactly when one is available.

Ensembles of random initializations estimate sensitivity to training choices under the fixed data and architecture. They do not automatically include waveform-model bias, simulation truncation error, omitted physics, or all out-of-distribution behavior. Calibrate any predictive interval against truly held-out high-fidelity points and publish coverage by region, not just globally.

The environment record pins training and validation data, split manifest, preprocessing, basis, architecture, initialization, optimizer and schedule, seeds, loss weights, stopping rule, precision, software and accelerator versions, operating system or container, inference code, and timing protocol. Record compiler kernels and deterministic settings when bitwise repeatability is claimed.

Reproducing the calculation and recognizing failure

Failure modes include leakage through preprocessing, seed-sensitive conclusions, shortcut learning from alignment or padding, broken symmetry or constraints, low median but unacceptable worst-case error, and extrapolation disguised as interpolation. A speedup also fails as scientific evidence when compared at different accuracy, on undisclosed hardware, or without preprocessing and batching costs.

A clean reconstruction verifies data and environment; rebuilds the basis and splits; trains every seed and baseline; sweeps coverage, capacity, optimization tolerance, and precision; computes regional errors, physical residuals, timing, and uncertainty coverage; runs the boundary and leave-region-out counterexample; and emits actual thresholds and failures. The result is valid only for the fixed modes, parameter domain, representation, and data provenance.

Review is triggered by a new waveform catalog or parameter domain, a changed architecture or training procedure, an independent worst-case or constraint failure, a software or hardware change that alters numerical behavior, or evidence that a downstream inference is sensitive to surrogate error.

Check your understanding

A surrogate has median mismatch $2\times10^{-5}$ over 20,000 test waveforms, but a cluster near the upper-$\chi_1$ boundary has mismatch $2\times10^{-3}$. An application requires every in-domain waveform to remain below $10^{-3}$. Can the median justify deployment?

Solution

No. The stated requirement is worst-case, and the observed boundary cluster violates it. Report the location and mechanism, densify or redesign the model there, and repeat a predeclared boundary holdout across seeds. Until the maximum in-domain error meets the threshold—or the validated domain is explicitly narrowed—the surrogate is not usable for that application, however strong its median performance.

Continue with Physics-Informed Constraints and Symmetries.

References

• Field, Scott E., Chad R. Galley, Jan S. Hesthaven, Jason Kaye, and Manuel Tiglio. “Fast Prediction and Evaluation of Gravitational Waveforms Using Surrogate Models.” Physical Review X 4 (2014): 031006. arXiv:1308.3565. DOI.
• Khan, Sebastian, and Rhys Green. “Gravitational-Wave Surrogate Models Powered by Artificial Neural Networks.” Physical Review D 103 (2021): 064015. arXiv:2008.12932. DOI.