Engineers validate computer models one piece at a time. A pump simulation is checked against a bench test, a surrogate neural network is checked against a fluid solver, and the solver is checked against a wind tunnel. Then the pieces are coupled into a larger system that nobody has validated as a whole, and the separate validation numbers stop meaning anything. This proposal treats each validation result as a graded premise: a statement of the form "model A deviates from reference B by at most this much, in this physical sense." It then treats the error budget of the coupled system as something deduced from those premises by a sound and complete logic of costs. The output is a machine-checkable certificate. It either proves that a new combination of validated components stays within a tolerance, or it names which component's budget breaks the claim. The combination must be one that was never tested end to end.
The mathematical home for this is the Lawvere quantale, meaning the non-negative reals with infinity, where "combining" means adding costs and "better" means smaller. Relations valued in this structure are graded, directed distances between possibly different spaces. Examples are a sensor grid versus a simulated field, or a coarse mesh versus a fine one. The claim is specific. Validation claims can be written in the affine logic $\mathbb{L}^*$ of Propositional Logics for the Lawvere Quantale, which is a proof system whose deductions reduce to linear inequalities decided by Fourier–Motzkin elimination and justified by Farkas' Lemma. With physical discrepancies split into several channels and coupling gains measured from data, this gives system-level error bounds on held-out component combinations. Those bounds should be both valid and tighter than the scalar error-propagation bounds engineers use today. We test this on PDEBench and on the AhmedML/DrivAerML simulation-to-wind-tunnel chains.
Core Claim
Two things from prior work meet here, and neither has been applied to physical models. First, Interface simulation distances replaces yes/no refinement between discrete automata with a directed distance. It proves that composing both sides with a third component never increases that distance, $d(A \parallel C, B \parallel C) \le d(A,B)$. This paper motivates our goal, but its theorem also constrains us. For coupled partial differential equation (PDE) models, errors passed across an interface can be amplified. We therefore do not expect non-expansion to hold, and we will measure how often it fails. Second, PDEBench shows that a low root-mean-square error (RMSE) can hide broken conservation, wrong boundary values, and distorted spectra. It introduces separate conservation, boundary, and frequency errors (cRMSE, bRMSE, fRMSE), but it reports them only as a dashboard, with no rule for combining them. We reuse its metrics as the coordinates of a product quantale $[0,\infty]^k$. We then write down how each coordinate travels across a coupling.
The coupling rule is a directed, multi-channel inequality. Let $e_i \in [0,\infty]^k$ be component $i$'s measured discrepancy vector against its own reference, with one coordinate per physics check. Let $e_i^{\mathrm{sys}}$ be its discrepancy once it is embedded in the coupled system. Then
$$e_i^{\mathrm{sys}} \;\le\; e_i + \sum_{j \in N(i)} \Gamma_{ij}\, e_j^{\mathrm{sys}},$$
where $N(i)$ is the set of components sharing an interface with $i$. The matrix $\Gamma_{ij}$ holds non-negative gains saying how much of neighbor $j$'s error in each channel shows up in each channel of $i$. For example, boundary error in $j$ may become interior and conservation error in $i$. This inequality is an $\mathbb{L}^*$ theory, and a target claim such as "drag error at most $\varepsilon$" is an $\mathbb{L}^*$ consequence question. Deciding it is a linear feasibility problem, and the Farkas multipliers that witness it are the certificate. We are honest about the lineage here. With one channel and known gains, this is a small-gain argument from control theory. The new parts are three. The gains are estimated per physics channel from validation data. The discrepancies are directed, so a surrogate that under-predicts drag is treated differently from one that over-predicts it. And the deduction comes with a completeness guarantee from Bacci et al., so "not derivable" really means "not implied by these premises," not merely "our heuristic failed."
The second use of the quantale structure is composing validation chains. A surrogate is validated against computational fluid dynamics (CFD), and the CFD against an experiment. Composing Lawvere-valued relations gives $(R;S)(x,z) = \inf_y \big(R(x,y) + S(y,z)\big)$, where $y$ ranges over intermediate references. In AhmedML these are the coarse-to-extra-fine meshes. Taking the minimum over several references gives a sound bound that is never looser, and often tighter, than routing through one fixed reference.
Research Program
The first piece of work is measurement. For each component we compute its directed discrepancy vector against its reference: signed field error split into positive and negative parts, plus PDEBench's conservation, boundary, and spectral errors. Measurement noise enters as an additive term in the premise, estimated from repeated reference runs. Because the approximate completeness in Bacci et al. excludes infinite values, we clip unbounded discrepancies and flag them rather than propagating infinity.
The second piece is gain estimation. On a calibration set of coupled systems, we fit the smallest non-negative gains $\Gamma_{ij}$ such that the coupling inequality holds on all but a chosen fraction of cases. This is a linear program, and we add a quantile-style slack so that coverage on new combinations has a nominal target. The gains are indexed by component type and interface kind, not by specific pairs, so they can transfer to combinations never seen.
The third piece is the deduction engine. It normalizes $\mathbb{L}^*$ judgments into linear inequalities as in Bacci et al., solves the feasibility problem, and emits the Farkas certificate. The certificate is a small vector of multipliers that an auditor can verify in linear time, and its largest entries identify the binding component.
Baselines
The naive reference point is reporting component metrics and doing nothing. We take the worst component's discrepancy as the system's error. This matches current practice when validated pieces are assembled, and it is expected to under-cover badly whenever interfaces amplify error.
The second baseline is the additive, non-expansive rule: system error equals the sum of component errors, with every $\Gamma_{ij}=0$. This is the direct transplant of the compositionality theorem in Interface simulation distances. Running it tells us whether the automata result transfers to PDEs.
The strongest classical competitor is scalar Lipschitz error propagation. It uses one RMSE channel and one gain per interface, estimated on exactly the same calibration couplings with the same linear-program fit and coverage target. This isolates the key ingredient, the multi-channel directed product quantale. If our tightness does not beat it, the channel split and directedness buy nothing, and we will say so.
The strongest data-rich competitor is end-to-end split conformal calibration on the coupled system. It is given the same total number of reference solves, but it spends them on coupled runs rather than component runs. This baseline is an upper reference on held-in combinations and a fair competitor on held-out ones.
Two further ablations isolate the remaining ingredients. The first swaps directed for symmetric discrepancies. The second, in the chain experiment, uses a single intermediate reference instead of the minimum over mesh resolutions.
For the automotive surrogate we use DragSolver from DragSolver, a multi-scale transformer that reports $R^2 = 0.9712$ on DrivAerNet++ and uses Monte Carlo dropout for uncertainty. Its dropout intervals, recalibrated on the same data, are the uncertainty baseline for drag claims. For PDEBench surrogates, the file lists Fourier Neural Operator, U-Net, and physics-informed network baselines, but no per-dataset error figures are recorded: no reported result found on Lacuna for FNO RMSE on the PDEBench 2D datasets, so we retrain them ourselves under PDEBench's protocol.
Beating the baselines means two things on held-out combinations. First, empirical coverage must be at or above the nominal target, for example 90%. Second, the median ratio of bound to realized error must be meaningfully smaller than under scalar Lipschitz propagation; we set a factor of 2 in advance as the bar for "worth the apparatus."
Why not just…?
Why not just propagate errors with Lipschitz constants or a sensitivity analysis? This is the real threat, and it is the scalar baseline above. The method in this proposal is that analysis when there is one channel and no directedness. The claim is that the extra structure matters: boundary errors turn into conservation errors across interfaces, and one-sided safety claims need one-sided bounds. PDEBench's evidence that RMSE hides boundary and conservation failures suggests a single channel will be loose. The ablation settles the question either way.
Why not just calibrate the whole coupled system with conformal prediction or Bayesian calibration? When coupled runs exist, that works and needs no logic. Compositional reasoning matters exactly when a new combination is proposed and running it end to end is expensive. AhmedML and DrivAerML runs cost on the order of 48 hours on clusters, and wind-tunnel time is scarcer still. Bayesian parameter calibration, as in Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization, tunes parameters of one model but says nothing about untested couplings. Our held-out-combination split measures directly whether component certificates can replace end-to-end calibration, and at what data cost.
Why not just train with physics penalties, as physics-informed networks do? Penalties change the model. They do not certify it, and PDEBench reports that physics-informed networks often struggle on stiff and multi-scale problems. Penalty-trained surrogates simply become more components for our certificates to cover.
Why not just report PDEBench's separate metrics? A dashboard cannot answer "does this assembly meet a 2% drag tolerance." The deduction step is what turns the dashboard into a yes/no claim with a certificate.
Proposed Experiments
The first testbed builds decomposed PDE systems from PDEBench's 2D Darcy flow, diffusion-reaction, shallow-water, and compressible Navier–Stokes data. Each rectangular domain is split into two to four subdomains. Each subdomain is solved by a component drawn from a pool: Fourier Neural Operator, U-Net, physics-informed network, or the reference solver at reduced resolution. Components exchange interface values by overlapping-Schwarz iteration, and the ground truth is the full reference solver. Gains are fitted on one set of type combinations and tested on disjoint ones. We report three things: coverage, the ratio of bound to realized error per channel, and the violation rate of the non-expansion law.
The second testbed is a validation chain. On AhmedML, surrogates are validated against the DDES runs at each mesh resolution. The DDES runs themselves match the experiment's drag and lift within 2%, as reported in AhmedML. The composed relation gives a surrogate-versus-experiment bound. On DrivAerML we repeat the construction with the ±1.5 drag-count statistical floor as the noise term and the Pininfarina comparison as the last link, following DrivAerML. Because wind-tunnel data exist only at nominal configurations, this testbed checks soundness at those points. It cannot measure coverage across a design space, and we will say so.
Three checks decide the outcome. The proposal succeeds if held-out coverage meets its target and multi-channel directed bounds are at least twice as tight as scalar Lipschitz bounds at matched coverage. It fails on either of two findings. The first is that fitted gains do not transfer, meaning coverage collapses on held-out combinations; that would show validation cannot be modular without end-to-end data. The second is that scalar propagation matches our tightness; that would show the quantale structure adds auditability but no statistical value. A third finding is informative whichever way it goes: the measured violation rate of $d(A\parallel C, B\parallel C) \le d(A,B)$ in coupled PDEs, which determines whether automata-style compositionality can be carried over to physical models at all.
References
- Propositional Logics for the Lawvere Quantale
- Interface Simulation Distances
- PDEBench: An Extensive Benchmark for Scientific Machine Learning
- DragSolver: A Multi-Scale Transformer for Real-World Automotive Drag Coefficient Estimation
- Bayesian Optimization for High-Dimensional Coarse-Grained Model Parameterization: A Case Study on Pebax Polymer
- AhmedML: High-Fidelity Computational Fluid Dynamics Dataset for Incompressible, Low-Speed Bluff Body Aerodynamics
- DrivAerML: High-Fidelity Computational Fluid Dynamics Dataset for Road-Car External Aerodynamics