Nuclear Data Adjustment for Nonlinear Applications in the OECD/NEA WPNCS SG14 Benchmark -- A Bayesian Inverse UQ-based Approach for Data Assimilation

This paper introduces a Bayesian Inverse Uncertainty Quantification (IUQ) approach for nuclear data adjustment within the OECD/NEA WPNCS SG14 benchmark, demonstrating its superior ability to replicate nonlinear model responses compared to traditional GLLS and MOCABA methods while validating the utility of low-correlation experiments for data assimilation.

The Gist

The Big Picture: Tuning the Nuclear Engine

Imagine you are trying to build a perfect model of a nuclear reactor. To do this, you need to know the exact "ingredients" (nuclear data) that make the reactor work. But, just like a recipe, your list of ingredients isn't perfect. There are small errors in the measurements of how fuel behaves, how neutrons bounce off atoms, and so on.

If your ingredients are slightly off, your model might predict the reactor will be safe when it's actually dangerous, or vice versa. Nuclear Data Adjustment is the process of fixing those ingredient lists by comparing your model against real-world experiments.

This paper is about a "cooking competition" (a benchmark) organized by the OECD to see which method is best at fixing these recipes, especially when the cooking gets complicated (nonlinear).

The Three Chefs (The Methods)

The paper tests three different chefs (methods) to see who can adjust the recipe best:

1. Chef GLLS (The Linear Calculator):

   • How they work: This chef assumes everything is a straight line. If you add a little more salt, the soup gets a little saltier in a perfectly predictable way.
   • The Problem: Real life isn't always a straight line. Sometimes, adding a little more salt makes the soup explode (nonlinear behavior). Chef GLLS gets confused when things get messy and can't predict the outcome accurately.
   • Verdict: Great for simple, predictable tasks, but fails when the physics gets complex.

2. Chef MOCABA (The Smart Sampler):

   • How they work: Instead of assuming a straight line, this chef tastes the soup thousands of times with slightly different ingredients to see what happens. They use a clever trick to turn those messy tastes back into a neat prediction.
   • The Verdict: Much better than Chef GLLS. They can handle the "explosive" nonlinear situations and give a good guess, though they might lose a tiny bit of detail in the process.

3. Chef IUQ (The Bayesian Detective):

   • How they work: This chef uses a powerful computer detective technique called Bayesian Inverse Uncertainty Quantification. Instead of guessing, they start with a "suspect list" (prior knowledge) and use real-world evidence (experiments) to narrow down the list of possible ingredient amounts. They don't assume anything is a straight line; they let the data speak for itself.
   • The Verdict: The most accurate chef. They capture the full, messy reality of the situation, including weird shapes and curves in the data that the other chefs miss. The downside? They are very slow and require a lot of computing power.

The Mystery of the "Unrelated" Clue

One of the most interesting parts of the paper is a lesson about Correlation.

In the nuclear world, scientists usually look for experiments that are "similar" to the reactor they are trying to predict. If an experiment looks 99% like the reactor, they use it. If it looks very different (low correlation), they usually ignore it.

The Paper's Discovery:
The authors found that sometimes, an experiment that looks totally different (low correlation) is actually super helpful.

• The Analogy: Imagine you are trying to figure out how a car engine works. You have a test drive on a flat highway (Experiment A) and a test drive on a steep, rocky mountain (Experiment B).
  • The highway drive looks very similar to your daily commute (High Correlation).
  • The mountain drive looks nothing like your commute (Low Correlation).
  • Old Thinking: "Ignore the mountain drive; it's too different."
  • New Thinking: "Wait! The mountain drive tests the engine under extreme stress that the highway never does. Even though it looks different, it teaches us something the highway drive couldn't."

The paper shows that by looking at how sensitive the models are to specific ingredients (sensitivity profiles) rather than just how similar they look, you can find hidden value in experiments that seem unrelated.

The Results: Who Won?

• For Simple, Straight-Line Problems: All three chefs did a good job. Chef GLLS was fast and accurate.
• For Complex, Wiggly (Nonlinear) Problems:
  • Chef GLLS failed. Their straight-line predictions didn't match the messy reality.
  • Chef MOCABA did well, getting close to the truth.
  • Chef IUQ was the winner. They perfectly matched the complex reality because they didn't force the data into a straight line.

The Takeaway

This paper tells us that while our old, simple tools (GLLS) are great for routine safety checks, we need smarter, more flexible tools (like Bayesian IUQ) to handle the complex, next-generation nuclear reactors of the future.

It also teaches us not to throw away "weird" experiments just because they don't look like the reactor we are studying. Sometimes, the most different-looking clues hold the key to solving the puzzle.

In short: To build safer, better nuclear models, we need to stop assuming everything is a straight line and start embracing the messy, complex reality of how the world actually works.

Technical Summary

1. Problem Definition

The paper addresses the challenge of nuclear data adjustment (calibrating nuclear cross-sections using experimental data) for applications that exhibit nonlinear behaviors.

• Context: The Organization for Economic Cooperation and Development (OECD) Nuclear Energy Agency (NEA) Working Party on Nuclear Criticality Safety (WPNCS) Sub-Group 14 (SG14) established a benchmark to test data adjustment techniques.
• The Benchmark: The problem involves 4 Experiments (Albert, Bohr, Chadwick, Dyson) and 3 Applications (Bravo, Castle, Trinity). The goal is to adjust 5 input nuclear cross-section parameters ($\theta$) to minimize the discrepancy between model predictions and experimental measurements, then predict the behavior of the Applications with uncertainty.
• The Challenge:
  • Some Applications (Bravo and Trinity) exhibit nonlinear responses over the parameter domain.
  • Some Experiments have low or negative correlation with the Applications, challenging traditional selection criteria based on correlation coefficients.
  • Hidden error sources are embedded in the models to test the robustness of the adjustment techniques.
• Objective: To evaluate the performance of Bayesian Inverse Uncertainty Quantification (IUQ) against traditional methods (Generalized Linear Least Squares - GLLS, and Monte Carlo Bayes - MOCABA) specifically in the context of nonlinear applications.

2. Methodologies Compared

The study compares three distinct approaches to data assimilation:

A. Generalized Linear Least Squares (GLLS)

• Mechanism: A traditional, deterministic method widely used in the nuclear industry (e.g., in SCALE and MCNP/Whisper). It assumes the model response is linear over the domain of uncertain parameters and that distributions are Gaussian (normal).
• Limitation: It relies on first-order Taylor expansions. It fails to capture higher-order moments (skewness, kurtosis) and produces erroneous predictions when the underlying physics is nonlinear.

B. Monte Carlo Bayes (MOCABA)

• Mechanism: A hybrid approach that avoids strict linearity assumptions by sampling from the prior parameter distribution to generate model responses. It uses an empirical quantile transform to map non-normal sampled responses to a normal distribution, performs the Bayesian update, and then inversely transforms the results.
• Strength: It can capture skewed posterior distributions better than GLLS.
• Limitation: It still relies on distribution transformations which can introduce slight distortions and requires careful handling of measurement error covariance during the transformation.

C. Bayesian Inverse Uncertainty Quantification (IUQ)

• Mechanism: A fully probabilistic approach using Markov Chain Monte Carlo (MCMC) to sample directly from the posterior parameter distribution.
• Key Features:
  • Model Discrepancy ($\delta$): Explicitly accounts for model bias (the difference between the model and reality) by treating it as a constant offset for each experiment in this specific benchmark.
  • Surrogate Modeling: Uses Gaussian Process surrogates to emulate the computationally expensive nuclear models, allowing for efficient MCMC sampling.
  • Direct Empiricism: The posterior predictive distribution is generated by directly running the model with posterior parameter samples, preserving the true shape (nonlinearity) of the response without linearization or transformation artifacts.

3. Key Contributions

1. Application of IUQ to Nuclear Data: The paper successfully adapts the Bayesian IUQ framework (previously used in thermal-hydraulics) to nuclear cross-section adjustment, demonstrating its viability for this domain.
2. Evaluation of Nonlinearity: It provides a rigorous comparison showing that while GLLS works for linear systems, it fails to replicate the true distribution of nonlinear applications. IUQ and MOCABA are shown to be superior for these cases.
3. Redefining Experimental Relevance: The study challenges the conventional reliance on the Pearson correlation coefficient ($c_k$) for selecting experiments. It demonstrates that:
   • Low correlation does not necessarily mean an experiment is uninformative.
   • Sensitivity profiles are a more robust metric for determining experimental relevance.
   • Experiments with low prior correlation (like "Chadwick") can significantly reduce posterior uncertainty if they provide unique information within the adjusted parameter subspace.
4. Impact of Model Discrepancy: The authors analyze the effect of including a model bias term ($\delta$), showing that it shifts the center of the posterior distribution and can alter the variance of the predictive distribution, sometimes reducing it by shifting the prediction to a region of lower model sensitivity.

4. Key Results

• Linear Applications (e.g., Castle): All three methods (GLLS, MOCABA, IUQ) produced consistent results, validating GLLS for linear criticality safety scenarios.
• Nonlinear Applications (e.g., Bravo, Trinity):
  • GLLS: Failed to capture the true distribution. It produced symmetric, Gaussian predictions that did not match the skewed, non-normal distributions of the computed model responses.
  • MOCABA: Successfully captured the skewness of the nonlinear responses, showing near-agreement with the computed model responses, though minor distortions from the transformation process were observed.
  • IUQ: Produced the most accurate representation of the posterior predictive distributions, as they were derived directly from the model evaluations without linearization or transformation.
• Case Studies (Experiment Selection):
  • Case 1 (Albert only): Significantly reduced uncertainty for highly dependent models (Bohr, Trinity).
  • Case 3 (Albert + Chadwick): Despite "Chadwick" having a low/negative correlation with Applications in the prior, its inclusion drastically reduced uncertainty. This occurred because Chadwick's sensitivity profile was distinct and remained informative even after the parameter space was constrained by Albert's data.
  • Case 5 (All Experiments): Adding highly correlated experiments (like Bohr and Dyson to Albert) provided diminishing returns in variance reduction but increased the robustness of the adjustment against measurement errors.

5. Significance and Conclusion

• Validation of IUQ: The study confirms that Bayesian IUQ is a powerful tool for nuclear data adjustment, particularly for nonlinear applications where traditional linear methods (GLLS) are insufficient.
• Methodological Shift: The paper argues for moving beyond simple correlation coefficients when selecting experiments for data assimilation. Instead, sensitivity analysis and the examination of sensitivity profiles within the posterior parameter space are critical for identifying informative experiments.
• Future Outlook: While IUQ offers superior accuracy for nonlinear problems, the authors note its higher computational cost compared to GLLS and MOCABA. They suggest that surrogate modeling is essential to make IUQ scalable for high-dimensional parameter spaces in future nuclear data evaluations.

In summary, this work demonstrates that for next-generation nuclear applications involving complex, nonlinear physics, Bayesian IUQ provides a more rigorous and accurate framework for data assimilation than traditional linear methods, provided computational costs are managed via surrogate modeling.