Learning Complex Physical Regimes via Coverage-oriented Uncertainty Quantification: An application to the Critical Heat Flux

This paper demonstrates that coverage-oriented uncertainty quantification methods, which integrate uncertainty directly into the optimization process, outperform post-hoc techniques in modeling the complex, multi-regime physical behaviors of Critical Heat Flux by producing models with both high predictive accuracy and physically consistent uncertainty estimates.

The Gist

Imagine you are trying to teach a robot how to predict when a pot of water will start boiling over violently (a phenomenon scientists call Critical Heat Flux or CHF). This isn't just a simple "boil at 100°C" situation. In a nuclear reactor, the water is under immense pressure, flowing at high speeds, and the physics changes drastically depending on exactly how fast it's moving and how hot it is.

Sometimes the water bubbles gently (Regime A), and sometimes it suddenly dries out and the metal gets dangerously hot (Regime B). The transition between these two states is chaotic and unpredictable.

Here is the problem: Standard AI models are like strict accountants. They look at all the data, calculate an average, and give you one single number: "The boiling point will be 100°C." If the real answer is 90°C or 110°C, the accountant is wrong. Worse, the accountant is confident they are right, even when they are guessing.

This paper proposes a new way to teach the AI. Instead of just being an accountant, the AI should become a weather forecaster.

The Core Idea: "Don't Just Guess, Know Your Confidence"

The authors argue that in complex physics, uncertainty isn't a mistake; it's a feature.

Think of it like driving a car:

• Standard AI: Tells you, "You will arrive in 30 minutes." (It doesn't care if it's raining or sunny).
• This New Approach: Tells you, "You will arrive in 30 minutes, but if you are in the rain, it might take 45. If you are on a highway, it might take 20."

The paper tests three different ways to teach the AI to be this kind of "weather forecaster" for nuclear reactors.

The Three Methods Tested

1. The "Post-It Note" Method (Conformal Prediction)

Imagine you train a strict accountant first. Once they are done, you go back and stick a Post-It note on their report that says, "Add a 20% safety margin just in case."

• Pros: It's safe. It guarantees that the real answer is usually within that margin.
• Cons: The accountant didn't learn anything new. They still think the world is simple. The "Post-It" is just a band-aid. It doesn't help the accountant understand why the rain makes the trip longer.

2. The "Dual-Brain" Method (Heteroscedastic Regression)

Instead of training the accountant and then adding a note, you train the AI with two brains at the same time.

• Brain 1: Predicts the boiling point.
• Brain 2: Predicts how "nervous" Brain 1 should be.
• How it works: As the AI learns, Brain 2 realizes, "Hey, when the water is moving fast, things get chaotic! I need to tell Brain 1 to be less confident and give a wider range of answers."
• Result: The AI learns that the rules of the game change depending on the situation. It doesn't just guess; it understands the physics of the chaos.

3. The "Safety Net" Method (Quality-Driven Learning)

This is similar to the Dual-Brain, but with a specific rule: "You must catch at least 95% of the real answers in your safety net, and you want that net to be as small (tight) as possible."

• The Magic: The AI is forced to stretch its safety net wide only when the physics gets weird (like during the transition from gentle bubbles to dryout) and keep it tight when things are calm. It learns to shape its own uncertainty to match the reality of the reactor.

The Big Discovery: The AI Learned the "Secret Language" of Physics

The most exciting part of the paper is what happened when they looked at the results.

The AI wasn't just giving better numbers; it was discovering physical truths on its own.

• The researchers found that the AI's "nervousness" (uncertainty) spiked exactly when the water was transitioning from one physical state to another (from bubbling to drying out).
• The AI didn't need a human to say, "Hey, watch out for the transition zone!" It figured it out by realizing, "I can't predict this part as well as the others, so I'll widen my safety net."

It's like teaching a dog to find a lost ball.

• Old way: You tell the dog, "Go to the park." The dog runs there, but if the ball is in the bushes, the dog might miss it.
• New way: You teach the dog, "If you smell the bushes, slow down and sniff carefully. If you smell the open grass, run fast." The dog learns to adapt its behavior to the terrain.

Why This Matters

In nuclear engineering, being "wrong" can be dangerous.

• If you are overconfident (think you know the answer when you don't), you might miss a safety hazard.
• If you are too cautious (give a huge safety margin), you might shut down a perfectly safe reactor, wasting energy and money.

This paper shows that by teaching AI to internalize uncertainty (to learn how to be unsure) rather than just calculating it afterwards, we get models that are:

1. Safer: They know when they are in a dangerous, chaotic zone.
2. Smarter: They learn the underlying physics of the system, not just the numbers.
3. More Efficient: They don't waste safety margins on calm days, saving resources.

The Bottom Line

The authors successfully showed that Uncertainty Quantification (UQ) shouldn't just be a safety check at the end of the process. It should be the teacher during the learning process.

By forcing the AI to learn how to be uncertain, the AI learns to see the world the way a physicist does: understanding that some situations are predictable, and others are wild, chaotic, and require a different kind of caution. It turns a "black box" calculator into a "self-diagnosing" expert.

Technical Summary

1. Problem Statement

The paper addresses a central challenge in Scientific Machine Learning (SciML): modeling physical systems governed by complex, multi-regime behaviors where data exhibits varying degrees of stochasticity (heteroscedasticity).

• The Specific Case: The study focuses on predicting the Critical Heat Flux (CHF), a phenomenon critical to nuclear reactor safety. CHF is characterized by highly non-linear dependencies on input parameters and distinct microscopic physical regimes: Departure from Nucleate Boiling (DNB) and Dryout.
• The Limitation of Standard ML: Traditional regression approaches prioritize minimizing global error metrics (e.g., Mean Squared Error). This often leads to models that suppress intrinsic data variability, resulting in overconfident predictions in turbulent or transition regimes. Standard Uncertainty Quantification (UQ) is often treated as a post-hoc safety check rather than an integral part of the learning process.
• The Goal: To develop models that not only predict CHF accurately but also internalize the underlying physical regimes by learning the heteroscedastic nature of the data, ensuring that uncertainty estimates are physically consistent and statistically calibrated.

2. Methodology

The authors propose a comparative analysis of UQ methodologies, contrasting post-hoc calibration against end-to-end coverage-oriented learning.

A. Dataset and Preprocessing

• Source: The U.S. Nuclear Regulatory Commission (NRC) dataset (24,579 entries from 60 sources).
• Filtering: To ensure physical consistency, the authors filtered the data to remove:
  • Entries with negative inlet subcooling (indicating flow instability).
  • Data from unreliable sources (e.g., imprecise collection methods or tubes too short for stable flow).
  • The final filtered dataset contains 22,872 points.
• Inputs: Geometric parameters ($D, L$) and physical properties ($P, G, T_{in}$). Note: Outlet vapor quality ($X$) is derived but not used as a direct input to prevent data leakage, though it is analyzed as a key physical indicator.

B. Baseline Architecture

• Residual Networks (ResNet): A deep learning architecture using fully connected layers, batch normalization, ReLU activations, and a Softplus output activation to ensure strictly positive CHF predictions.
• Optimization: Trained using Adam/AdamW with early stopping, minimizing Root Mean Squared Percentage Error (RMSPE).

C. Uncertainty Quantification Approaches

The study compares four distinct strategies:

1. Conformal Prediction (CP) - Post-hoc:

   • Vanilla CP: Calculates constant uncertainty bounds based on residuals from a calibration set.
   • Adaptive CP: Uses a score function (residuals normalized by a heteroscedasticity estimator) to create variable bounds. This method calibrates a frozen model without altering its internal representation.

2. Heteroscedastic Regression (HR) - End-to-End:

   • The model is trained to output both the mean ($\mu$) and the standard deviation ($\sigma$) simultaneously.
   • Loss Function: Generalized Negative Log-Likelihood (NLL) loss.
   • Variants: Single-task, Multi-Task Learning (MTL), and Transfer Learning (TL) from a pre-trained regression model.

3. Quality-Driven (QD) Prediction - End-to-End:

   • The model outputs the prediction and explicit lower/upper bounds ($y_L, y_U$).
   • Loss Function: A composite loss minimizing RMSPE and the average interval width, subject to a coverage constraint (Penalty if the Prediction Interval Coverage Probability < target $\alpha$).
   • Allows for asymmetric bounds, capturing potential differences between underestimation and overestimation.

4. Bayesian Heteroscedastic Regression (BHR) - End-to-End:

   • Implements HR using a Bayesian Neural Network (BNN) via Variational Inference.
   • Goal: To explicitly separate aleatoric uncertainty (data noise) from epistemic uncertainty (model ignorance).
   • Loss: $\beta$-Negative Evidence Lower Bound (ELBO).

3. Key Contributions

• Active vs. Passive UQ: The paper demonstrates that treating UQ as an active component of the optimization process (End-to-End) forces the model to learn a representation that aligns with physical regimes, whereas post-hoc methods (CP) only adjust bounds around a static representation.
• Physical Regime Discovery: The models successfully identified the transition between DNB and Dryout regimes autonomously. This was evidenced by a localized spike in uncertainty bounds at specific outlet quality values ($X \in [0.25, 0.5]$), matching physical expectations without explicit regime labeling.
• First Application of QD Loss to CHF: This work introduces Quality-Driven loss functions to the CHF regression problem, showing their efficacy in achieving coverage-based learning.
• Aleatoric Dominance: Bayesian analysis confirmed that the total uncertainty in CHF prediction is dominated by aleatoric uncertainty (intrinsic physics variability) rather than epistemic uncertainty (model lack of knowledge), validating the use of heteroscedastic models.

4. Results

The methods were evaluated on the NRC dataset using predictive metrics (RMSPE, MAPE) and UQ metrics (PICP, Informativeness, Calibration, UQF-score).

• Predictive Accuracy: All methods achieved an RMSPE of approximately 10-11%, comparable to or better than the state-of-the-art (Groeneveld LUT and previous ML benchmarks).
• Coverage (PICP):
  • Adaptive CP: Achieved high coverage (>95%) but with wider, less informative bounds in some regimes.
  • End-to-End (HR, QD, BHR): Successfully met the 95% target coverage while maintaining tighter bounds in stable regimes.
• Physical Consistency (The "Regime" Effect):
  • All heteroscedastic models (HR, QD, BHR) and Adaptive CP showed a clear correlation between uncertainty width and Outlet Quality ($X$).
  • Low $X$ (DNB): Low uncertainty.
  • Transition ($X \approx 0.25-0.5$): Significant spike in uncertainty width (up to 100% relative width), correctly flagging the physical instability.
  • High $X$ (Dryout): Moderate uncertainty.
• Comparison of Approaches:
  • CP (Post-hoc): Good for safety guarantees but does not improve the model's internal understanding of physics.
  • HR & BHR: Provided a balanced trade-off between accuracy and uncertainty. BHR confirmed that model uncertainty is negligible compared to data noise.
  • QD: Offered the unique ability to learn asymmetric bounds, revealing that the model tends to overestimate CHF more than underestimate it in certain regimes.
  • Transfer Learning (TL): Proved effective in reducing training costs while maintaining high calibration, though it sometimes resulted in wider bounds (conservative estimates).

5. Significance and Conclusion

The paper argues for a paradigm shift in Scientific ML: Uncertainty Quantification should be an active learning objective, not a post-processing step.

• Scientific Impact: By integrating coverage-oriented objectives directly into the loss function, models become "self-diagnosing" tools. They do not just predict a value; they learn where the physics is complex and where the model's confidence should naturally decrease.
• Safety Implications: In high-stakes fields like nuclear engineering (governed by Best Estimate Plus Uncertainty, BEPU, frameworks), this approach ensures that safety margins are not just statistically valid but physically consistent. The model learns to flag transition zones (like the DNB-to-Dryout shift) automatically.
• Generalizability: The methodology is applicable to any multi-regime physical system where data heteroscedasticity is a signature of underlying physical transitions.

In summary, the study proves that coverage-oriented learning yields models that are not only more accurate but also more interpretable and reliable, as they inherently capture the stochastic nature of complex physical phenomena.