Two-Step MV-DeepONet: Probabilistic Operator Learning for Uncertainty Propagation Driven by Random Input Fields

Based on the paper, here is a detailed summary:

This paper develops a two-step MV-DeepONet framework for forward uncertainty propagation driven by random input fields, with a focus on recovering the total predictive covariance of high-dimensional output fields. The total predictive covariance is decomposed into the covariance of conditional means across input realizations and the average conditional predictive covariance. The paper identifies that Probabilistic DeepONet (Prob-DeepONet) provides lightweight uncertainty quantification by predicting pointwise Gaussian means and variances in a single forward pass, but its conditional predictive covariance is restricted to a diagonal form. To address this limitation, the proposed framework incorporates two principal modifications: (1) two-step training is used to decouple output-basis learning from the input-to-coefficient mapping, together with basis orthogonalization and subspace rotation; (2) Gaussian probabilistic modeling is transferred from the high-dimensional physical output space to the low-dimensional rotated coefficient space. Mapping these probabilistic coefficients through the shared basis induces a generally non-diagonal conditional predictive covariance in the physical output space while retaining single-pass inference.

The paper establishes a theoretical analysis of output covariance recovery through a Frobenius-norm error decomposition and a corresponding upper bound. The analysis identifies four principal factors governing covariance recovery: low-rank covariance compressibility, trunk-subspace approximation, finite-sample statistical error, and coefficient-space covariance estimation. The resulting low-rank formulation avoids explicit learning or storage of the full high-dimensional covariance matrix while preserving lightweight, single-pass inference.

Numerical experiments are conducted on three representative problems governed by partial differential equations (reaction–diffusion, Burgers, and Darcy equations) and a hypersonic blunt-body aerothermal problem. The results show that compared with Prob-DeepONet, the proposed method exhibits improved generalization to unseen inputs, produces tighter and more structured uncertainty bands, and recovers correlation patterns consistent with the underlying physical mechanisms. Conformal calibration further improves empirical interval coverage, while two-step MV-DeepONet maintains sharper prediction intervals.

The paper concludes that the two-step strategy helps reduce the mutual compensation between basis functions and coefficients, while coefficient-space probabilistic modeling propagates uncertainty jointly to multiple output locations through the shared basis functions. Future work may investigate adaptive mode-selection strategies and non-Gaussian coefficient-space distributions for cases where conditional predictive distributions exhibit non-Gaussian features.

Based on this paper, here are the specific improvements I can make to AI systems, and what the improved system can do:

1. Improve uncertainty quantification in neural operators for PDEs

• Improvement: Replace diagonal-only covariance predictions (as in Prob-DeepONet) with a two-step training that learns a shared, orthogonalized basis and then models probabilistic coefficients in a low-dimensional rotated space. This induces a full, non-diagonal covariance in the output space without extra inference cost.

• What the improved system can do: For a given random input field (e.g., uncertain material properties, boundary conditions), the AI can output not just a mean prediction but a structured, spatially correlated uncertainty band (e.g., for temperature or velocity fields). It will correctly indicate that errors at nearby points are correlated, not independent, leading to more realistic confidence intervals for engineering design or risk assessment.

2. Enable single-pass, low-rank covariance recovery for high-dimensional outputs

• Improvement: Use the Frobenius-norm error decomposition to guide training—explicitly minimize the four identified error sources (low-rank compressibility, trunk-subspace approximation, finite-sample error, coefficient covariance estimation). This avoids storing or computing the full N×N covariance matrix.

• What the improved system can do: For outputs with millions of spatial degrees of freedom (e.g., 3D aerothermal fields), the AI can produce a full predictive covariance in a single forward pass, using only a small number of latent coefficients. This makes real-time uncertainty propagation feasible for large-scale simulations, which was previously computationally prohibitive.

3. Improve generalization to unseen input distributions

• Improvement: Decouple basis learning from input-to-coefficient mapping via two-step training, and add subspace rotation to reduce mutual compensation between basis and coefficients.

• What the improved system can do: When given a new input field that is outside the training distribution (e.g., a different roughness pattern on a hypersonic vehicle), the AI will produce more accurate mean predictions and more reliable uncertainty estimates, rather than overconfident or poorly calibrated intervals. This is critical for extrapolation tasks in scientific discovery or design optimization.

4. Enable joint propagation of uncertainty across multiple output locations

• Improvement: Model probabilistic coefficients in the rotated latent space, then map them through the shared basis to the physical space. This couples the uncertainty at all output points through the shared latent variables.

• What the improved system can do: For a stochastic PDE solver, the AI can answer queries like What is the probability that the temperature exceeds a threshold at both point A and point B simultaneously?—something a diagonal-covariance model cannot do. This enables more accurate reliability analysis and failure probability estimation for coupled physical systems.

5. Provide calibrated, sharper prediction intervals via conformal calibration

• Improvement: Integrate the proposed covariance structure with conformal prediction to adjust interval widths based on empirical coverage, while preserving the non-diagonal correlation structure.

• What the improved system can do: The AI will output prediction intervals that meet a user-specified coverage level (e.g., 95%) on new data, but with narrower intervals than a diagonal model, because it correctly accounts for spatial correlations. This reduces wasted safety margins in engineering applications (e.g., less over-conservative design of thermal protection systems).

6. Support adaptive mode selection for covariance compression

• Improvement: Use the theoretical error bound to dynamically select the number of retained basis modes based on the desired accuracy vs. computational cost trade-off.

• What the improved system can do: For a given problem, the AI can automatically decide how many latent coefficients to use—fewer for smooth fields (fast, low memory) and more for highly heterogeneous fields (accurate, higher cost)—without manual tuning. This makes the system adaptable to different PDEs and input random fields.