A conformal prediction framework for uncertainty quantification in physics-informed neural networks

Yifan Yu, Cheuk Hin Ho, Yangshuai Wang

TL;DR

This paper introduces a novel framework that uses conformal prediction to provide statistically guaranteed uncertainty estimates for Physics-Informed Neural Networks (PINNs). This approach addresses the critical lack of reliable uncertainty quantification in PINNs, enabling more trustworthy predictions for complex physical systems.

What Problem Does This Paper Address?

Physics-Informed Neural Networks (PINNs) are powerful tools for solving partial differential equations (PDEs) by integrating physical laws into their training. However, existing methods for quantifying the uncertainty of PINN predictions often lack rigorous statistical guarantees. This means we can't be sure how reliable the predicted uncertainty is, which is a major hurdle for using PINNs in critical scientific and engineering applications where understanding prediction confidence is paramount. The paper aims to bridge this gap by providing a statistically principled, distribution-free framework for uncertainty quantification in PINNs.

Methods

The core contribution is the integration of Conformal Prediction (CP) with PINNs. CP is a statistical framework that provides prediction intervals with guaranteed finite-sample coverage, meaning the true solution will fall within the predicted interval a specified percentage of the time, regardless of the underlying data distribution.

Here's how it works:

  • Distribution-Free Uncertainty: The framework is "distribution-free," meaning it doesn't assume any specific probability distribution for the data or model errors. This makes it more robust.
  • Post-Hoc Calibration: CP acts as a "post-hoc wrapper." It takes the raw uncertainty estimates from existing PINN methods (like dropout, Bayesian inference, or geometric distance-based heuristics) and calibrates them to produce statistically valid prediction intervals.
  • Nonconformity Scores: The method works by calculating "nonconformity scores" on a separate calibration dataset. These scores measure how "unusual" a data point is compared to the model's prediction.
  • Local Conformal Prediction (Local CP): To handle situations where the uncertainty varies spatially (heteroskedasticity), the authors propose an extension called "Local CP." This method adapts the prediction intervals to local noise levels, leading to tighter and more informative intervals in regions with less uncertainty and wider intervals where uncertainty is higher. This is achieved by learning an input-dependent conditional quantile function.

Datasets

The paper uses standard benchmark PDEs for evaluation, not new datasets:

  • 1D Poisson equation
  • 2D Allen-Cahn equation
  • 3D Helmholtz equation
  • 1D Damped Harmonic Oscillator (used for testing local CP's adaptivity to heteroskedasticity)

For each PDE, synthetic data is generated based on the exact solution, with added noise to simulate real-world measurement errors. A portion of this synthetic data is used for training, and a separate portion for calibration.

Key Insights and Ideas

  • Bridging Deterministic PINNs and Rigorous UQ: The most significant insight is that a statistically principled, distribution-free UQ method like CP can be seamlessly integrated with existing, often deterministic, PINN architectures. This unlocks reliable uncertainty quantification without requiring a full probabilistic re-formulation of PINNs.
  • CP as a Universal Calibrator: The paper demonstrates that CP can effectively calibrate a variety of heuristic UQ methods, transforming potentially over-confident or under-confident estimates into reliable prediction intervals.
  • Local CP for Heteroskedasticity: The introduction of Local CP is a key advancement. It shows that adapting uncertainty quantification to spatially varying noise levels is crucial for accurate and informative predictions in many real-world PDE problems. Standard CP can be overly conservative in some regions and under-confident in others when noise is not uniform.
  • No Retraining Needed: A major advantage is that CP and Local CP are post-hoc methods. They calibrate existing PINN predictions without requiring any retraining of the underlying neural network, making them computationally efficient to apply.

Key Results

  • Reliable Calibration: Across all tested PDEs and heuristic UQ methods, the proposed CP framework consistently produces prediction intervals with empirical coverage close to the target confidence level. For example, in the 1D Poisson problem, naïve B-PINN intervals substantially under-covered the truth, while after CP, they achieved near-expected coverage.
  • Improved ACD: The Average Coverage Deviation (ACD), a metric for overall calibration quality across different significance levels, was significantly reduced by CP and Local CP compared to uncalibrated methods. For instance, in the 2D Allen-Cahn problem, ACD for the Geometric Distance method dropped from 0.3395 to 0.0329 after CP.
  • Local CP Adapts to Noise: In experiments with heteroskedastic noise (1D Damped Harmonic Oscillator and 2D Allen-Cahn with irregular noise patterns), Local CP demonstrated superior performance. It generated intervals that were tighter in low-noise regions and wider in high-noise regions, leading to better local coverage and sharpness compared to standard CP, which struggled with non-uniform noise. For example, in noisy regions of the 1D Damped Harmonic Oscillator, Local CP achieved 0.94 coverage compared to standard CP's 0.76.
  • Robustness: The framework proved robust across different PDE complexities and noise scenarios, consistently improving the reliability of uncertainty estimates.

Limitations and Future Work

  • Limitations:
    • The paper acknowledges that while CP guarantees finite-sample coverage, the intervals can sometimes be wider than desired, especially when the underlying heuristic UQ method is poorly calibrated. This is the trade-off for guaranteed reliability.
  • The performance of Local CP relies on the quality of the baseline heuristic uncertainty estimator (σ).
  • Future Work:
    • Extending the framework to inverse and partially observed problems.
  • Applying it to time-dependent PDEs.
  • Integrating with operator-learning paradigms like DeepONets and Fourier Neural Operators.
  • Exploring alternative nonconformity scores (e.g., energy-norm residuals, multi-output measures) for task-specific calibration.

Summary updated

BibTeX
@article{yu2026conformal,
  author = {Yifan Yu and Cheuk Hin Ho and Yangshuai Wang},
  title = {{A conformal prediction framework for uncertainty quantification in physics-informed neural networks}},
  journal = {Journal of Computational Physics},
  year = {2026},
  doi = {10.1016/j.jcp.2026.114979},
}
Figure