A new approach for the critical heat flux application - Michele CAZZOLA

Value of the Information & Strength of the Argument

The presentation provides valuable insights into applying modern machine learning techniques to a critical nuclear safety problem. The speaker systematically compares multiple models and uncertainty quantification methods, providing quantitative results (e.g., RMSPE values, coverage percentages). The argumentation is solid, with clear explanations of the methodologies and their motivations. The discussion of the physics-informed network’s degradation due to bias in the prior model is particularly insightful. The speaker also acknowledges limitations, such as the dataset’s age and the need for generalization to real geometries. Overall, the value is high for researchers in thermal hydraulics and machine learning, offering a practical framework for uncertainty quantification in safety-critical applications.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is generally good for a master’s thesis presentation. The methodology is clearly described, and results are presented with numerical metrics. However, the presentation lacks detailed citations to specific literature; only the OECD/NEA benchmark and the NRC dataset are mentioned. The title accurately reflects the content. The speaker does not provide external sources or references in the video description, so the sources cited are limited to those mentioned in the talk. The adequacy between title and content is high.

The title accurately reflects the content, which focuses on a new machine learning approach for critical heat flux prediction and uncertainty quantification.

7/10

The presentation is a master's thesis defense, describing original research with a clear methodology, quantitative results, and references to an international benchmark. However, it is not peer-reviewed, and the dataset and code are not publicly available in the video. The speaker acknowledges limitations and uncertainties.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and context of critical heat flux (CHF) in nuclear reactors.
• 02:00Challenges in CHF prediction: nonlinearity and two underlying phenomena (DNB and dryout).
• 04:00Description of the NRC dataset and the benchmark exercise.
• 06:00Research objectives: optimize CHF predictions and estimate uncertainty bounds.
• 08:00Machine learning models tested: SVR, MLP, ResNet, and physics-informed network.
• 10:00Results of model comparison: best performance with 8-layer ResNet on filtered data.
• 12:00Introduction to uncertainty quantification: conformal prediction and quality-driven approach.
• 14:00Baseline conformal prediction results: constant bounds with 93% coverage.
• 16:00Adaptive conformal prediction: variable bounds with 97% coverage and 23% average relative width.
• 18:00Quality-driven approach: asymmetric bounds, 96% coverage, and joint optimization with RMSPE.
• 20:00Comparison of uncertainty bounds across methods and analysis of input feature influence.
• 22:00Conclusions and future work: article in preparation, international benchmark.
• 24:00Q&A session: generalization to different geometries and data reliability.

• OECD/NEA benchmark on machine learning for CHF — Mentioned as the international activity framing the project.
• NRC dataset — Dataset released by the US Nuclear Regulatory Commission, used for training and evaluation.

• OECD/NEA benchmark on machine learning for CHF — The project aligns with the goals of this international benchmark.