Wrapping a learned physics simulator in conformal prediction produces calibrated error bars without assuming anything about the data distribution. Conformal prediction is a post-hoc recipe: measure how wrong the model was on held-out examples, then widen its outputs by the right quantile of those errors. The resulting bands, however, are built one grid cell or one station at a time. They happily contain fields with negative density, mass that appears from nowhere, or a dew point above the air temperature. That matters most when the band is used downstream. Someone who needs a bound on total mass in a region, or the net flux across a boundary, must add up per-cell intervals, and those sums are often enormous because the band never learned that the cells are tied together by a conservation law.
We propose to build the physics into the shape of the conformal set, so that only near-physical fields are included, while keeping the coverage guarantee even when the physics is wrong. The obstacle is that the physics we can write down is a parametric model that is almost always misspecified. Fitted diffusion coefficients are off, sources and boundary inflows go unmodeled, and approximate balances such as pressure–wind consistency only hold roughly. If we intersect the band with the exact constraint, it stops covering the truth as soon as the constraint is wrong. Our fix is to measure, on held-out data, how badly the true states violate the assumed physics, and to make that violation part of the conformal score. Misspecified physics then costs set size but never coverage, and a well-specified constraint makes the set sharply smaller.
Core Claim
Let $\hat y(x)$ be any forecaster's predicted field (or trajectory) given input $x$, and $\hat\sigma_j(x)$ a per-location scale. Let $g(y;\hat\theta)$ be a vector of physics residuals under a parametric physical model with parameters $\hat\theta$ fitted on the training split. Examples are a discrete mass budget between steps, nonnegativity, a discretized divergence, Kirchhoff-type node balances, or the inequality that dew point cannot exceed temperature. We score a calibration example by
$$ s(x,y)=\max\Big(\max_j \frac{|y_j-\hat y_j(x)|}{\hat\sigma_j(x)},\ \max_k \frac{|g_k(y;\hat\theta)|}{\tau_k}\Big), $$
where $\tau_k>0$ are weights chosen on a separate tuning split. The prediction set is $C(x)=\{y: s(x,y)\le \hat q\}$, where $\hat q$ is the usual split-conformal quantile of the calibration scores. Standard split-conformal reasoning gives $P(Y\in C(X))\ge 1-\alpha$ whenever calibration and test units are exchangeable. This holds no matter how wrong $\hat\theta$ or the physical model class is, because the true $Y$'s own physics residual is what gets calibrated.
The set is a box intersected with a slab around the assumed physics. When the constraints are linear, the tightest bound on any linear quantity of interest over $C(x)$ is a small linear program. Because all such bounds come from one set, they hold simultaneously for every downstream query at level $1-\alpha$.
The novel claims are three. Physics can tighten conformal sets with no loss of validity. Misspecification of the physics shows up only as a wider slab, and the slab can be switched off ($\tau_k=\infty$) by tuning-split selection, so the method should not do worse than ignoring physics beyond tuning noise. The gains concentrate on aggregate quantities that per-cell bands handle badly.
Relation to Prior Work
Two papers motivate the gap. A conformal prediction framework for uncertainty quantification in physics-informed neural networks wraps PINN uncertainty heuristics in split and locally adaptive conformal prediction and fixes their calibration. Its bands are still per-point intervals whose envelopes need not satisfy the PDE, and it assumes the operator is known. Robust Conformal Prediction under Distribution Shift via Physics-Informed Structural Causal Model puts reaction–diffusion and SIR dynamics into the point predictor to make conformal scores more stable across environments. Physics there shapes the center of the band, not its geometry, and the paper itself notes that it needs well-specified physical structure. We do the opposite: the predictor can be anything, and physics enters the set with a calibrated tolerance.
Delving into Non-Exchangeability for Conformal Prediction in Graph-Structured Multivariate Time Series (SCALE) shows that ignoring spatiotemporal coupling drops nominal 95% coverage to 70%. It repairs this with graph-wavelet decompositions. That work constrains our design: on single-realization time series, exchangeability is the binding problem, so our main validity claims are made where whole trajectories from independent initial conditions are the calibration units. SCALE is also a strong width baseline on graph data. Semiparametric conformal prediction builds joint multi-target sets by modeling score dependence with vine copulas. It shrinks sets by learning statistical dependence, whereas we shrink them with structural constraints known a priori, which also apply in thousands of dimensions where copulas are impractical.
The misspecification literature plays a specific role here. A view on model misspecification in uncertainty quantification argues that the gap between the best in-class model and reality must count as uncertainty. Our slab is a direct, measurable version of that gap for the physical model. The file's research summary suggests that Neyman-orthogonal or doubly robust scores, in the sense of GDR-learners or Estimating Identifiable Causal Effects through Double Machine Learning, are needed to stop physical nuisance errors from invalidating conformal coverage. We think this is mistaken for split conformal prediction. There, nuisance errors in $\hat\theta$ or $\hat y$ affect efficiency, not marginal validity, which rests only on exchangeability. Orthogonalization is the right tool for estimating $\theta$ itself, and we leave it aside.
On the physics side, Learned Simulators for Turbulence reports that constraint losses mitigate but do not stop long-run conservation drift. Semigroup Consistency as a Diagnostic for Learned Physics Simulators finds that semigroup error predicts rollout failure (Spearman ρ = 0.635), even though it fails as a training regularizer. Both results say that physics is more reliably used at inference than as a training penalty, and that is where we put it. We reuse semigroup error as a candidate scale $\hat\sigma_j(x)$.
Research Program
The first component is the construction above and its rollout version. For autoregressive forecasts, the score takes a maximum over lead times $h$ as well as locations, with lead-time-specific scales $\hat\sigma_{j,h}$. One quantile then gives a whole-trajectory set valid jointly across horizons. Calibration units are entire trajectories from independent initial conditions, which are exchangeable even though steps within a trajectory are not.
The second component is selecting $\tau$ and the physics model. We fit $\hat\theta$ on the training split. On a tuning split we choose $\tau_k$ from a grid that includes $\infty$, minimizing average quantity-of-interest width. The calibration split then sets $\hat q$. This three-way split keeps validity exact and lets the data decide how much to trust each constraint.
The third component is downstream bounds. For linear constraints (sums, nonnegativity, discrete divergence, node balances, linear inequalities between variables), bounds on linear functionals are linear programs. For one nonlinear case, energy that must not increase in a damped oscillator, we use the exact low-dimensional set. We also supply a sampler that draws members of $C(x)$ for ensemble-style use.
The fourth component is analysis. We will bound the expected box width as a function of the calibrated physics residual quantile. The prediction is that the gain on an aggregate functional shrinks roughly linearly as the slack grows, because the slab width is proportional to the slack, and vanishes once the slab stops binding. We will check this against a controlled misspecification sweep.
Baselines
The naive reference is a point forecaster with no uncertainty, plus plain split conformal with a constant residual quantile. Downstream bounds are obtained by summing per-cell intervals. This shows how bad unstructured bands are for aggregates.
The strongest per-location method is Local CP from A conformal prediction framework for uncertainty quantification in physics-informed neural networks. It reports 0.94 coverage in the noisy region of the damped oscillator, against 0.76 for standard CP, and cuts average coverage deviation on 2D Allen–Cahn from 0.3395 to 0.0329. We run it with the same forecaster, the same normalizer family and the same three-way data split. Our method must match its coverage and beat its width on aggregate quantities.
On graph time series, SCALE (code) is the strongest competitor. It reports a roughly 19.6% interval-width reduction over the next-best method on Electricity, and it is run from the released code with its default tuning budget. PI-SCM is a physics-in-the-predictor competitor, but no reported result was found on Lacuna for its coverage-divergence values. We will reimplement it on the traffic setting and report our own numbers.
For joint multi-output sets on small targets we include the copula method of Semiparametric conformal prediction, which reports an efficiency score of 3.5 against 4.1 for independent CP on the Penicillin simulator.
Three ablations isolate the key ingredient. The exact intersection ($\tau\to 0$ with the uncalibrated constraint imposed) tests whether calibrated slack is necessary. The union-bound version, with separate quantiles for data and physics terms at $1-\alpha+\beta$ and $1-\beta$, tests whether the joint max-score helps. Removing $\tau=\infty$ from the selection grid tests the "never worse" safeguard.
Beating the baselines means two things. Coverage must be within Monte Carlo error of $1-\alpha$ on independent test trajectories. Downstream bounds must be significantly narrower than those of Local CP and SCALE on aggregate functionals, with the comparison run across seeds at equal compute.
Why not just…?
Why not just conformalize the quantity of interest directly? For one aggregate fixed in advance, such as total domain mass, a dedicated split-conformal interval on that scalar is valid and probably as tight as ours. This is a real threat, and we expect to tie or lose there. The advantage of a set comes when queries are many or chosen afterwards: arbitrary regional masses, boundary fluxes, extremes over subregions. Per-query CP then needs fresh calibration and a multiplicity correction. We test this directly against per-query CP with Bonferroni correction over $K$ queries, varying $K$ from 1 to hundreds, and will report the crossover honestly.
Why not just project the forecast onto the constraint and then run standard conformal prediction? Projection or hard-constrained architectures fix the center of the band, not its shape. A box around a perfectly conserving forecast still contains fields that violate conservation, and its aggregate bounds are still sums of widths. Projection with exact physics also inherits misspecification bias. Under the physics most likely to be wrong, it can move the center away from the truth and widen the band. "Projected forecaster plus Local CP" is a baseline in every experiment.
Why not just add physics losses during training? Learned Simulators for Turbulence reports persistent drift despite constraint losses, and Semigroup Consistency as a Diagnostic for Learned Physics Simulators reports that consistency penalties fail to improve rollouts. In any case, training losses give no coverage statement. Our method is complementary. We apply it on top of physics-regularized forecasters too, and expect smaller but nonzero gains, since better forecasts shrink the box but do not tie its cells together.
Why not just intersect the conformal box with the exact physics? If the physical model is exactly right, this is valid and slightly tighter than our version. Under misspecification it silently loses coverage. The misspecification sweep below is built to show where this happens.
Proposed Experiments
The core testbed is PDE rollouts with controllable misspecification, where calibration trajectories are exactly exchangeable. We use 1D heat and viscous Burgers from the semigroup study, KS-1D and 2D decaying turbulence from the turbulence study, and the damped oscillator from the conformal PINN study. Forecasters include a Dil-ResNet-style simulator and a neural operator, each with and without physics losses.
The physics constraint is a discrete mass or energy budget computed under a parametric model. We degrade that model along controlled axes: wrong viscosity, an unmodeled source term, and periodic boundaries assumed where inflow exists. Metrics are joint trajectory coverage, coverage per lead time, linear-program bound width for families of regional mass and flux queries, and the fraction of sampled set members that violate the true physics. On these rollout benchmarks, no reported result was found on Lacuna for interval width or coverage. For the damped oscillator, the Local CP coverage numbers above are the reference.
The decisive check is the sweep. Plot coverage and aggregate width against misspecification severity for our method, the exact intersection and Local CP. Our method should hold nominal coverage throughout, be clearly narrower than Local CP at low severity, and converge to Local CP's width at high severity. The exact intersection should lose coverage as severity grows. If our width does not fall below Local CP even with correctly specified physics, the idea fails.
The real-data test is WEATHER-5K. It pairs one exact physical inequality (dew point ≤ temperature, wind speed ≥ 0) with one approximate balance (pressure–wind consistency, as used in the PhysicsFormer loss). This lets us see the slack collapse for exact constraints and stay wide for approximate ones. The forecaster is PhysicsFormer-style; that model reports overall temperature MAE of 2.34 in Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting. Because station series are not exchangeable, we calibrate on station-year blocks, test on later years, and report coverage per year and per region, including the sparse regions that benchmark flags. Any guarantee there is explicitly approximate.
As a stretch, on traffic data with flow channels we test node-balance constraints against SCALE and the reimplemented PI-SCM.
Success means exact coverage on the PDE suite, monotone width degradation under misspecification with a bounded worst case, and a clear win over per-query conformal prediction once more than a handful of aggregate queries are asked.
References
- A conformal prediction framework for uncertainty quantification in physics-informed neural networks
- Robust Conformal Prediction under Distribution Shift via Physics-Informed Structural Causal Model
- Delving into Non-Exchangeability for Conformal Prediction in Graph-Structured Multivariate Time Series
- Semiparametric conformal prediction
- A View on Model Misspecification in Uncertainty Quantification
- GDR-learners: Orthogonal Learning of Generative Models for Potential Outcomes
- Estimating Identifiable Causal Effects through Double Machine Learning
- Learned Simulators for Turbulence
- Semigroup Consistency as a Diagnostic for Learned Physics Simulators
- Benchmarking Physics-Informed Time-Series Models for Operational Global Station Weather Forecasting