Turbulent Combustion-From Governing Principles to ML-Enhanced Combustion Modelling, Parente Day5 Pt3

Value of the Information & Strength of the Argument

The lecture provides significant value by bridging traditional combustion modeling with modern machine learning techniques. The speaker presents a clear argument for combining physics-based models with data-driven methods to overcome computational bottlenecks. The discussion of chemical reactor networks and their enhancement with graph neural networks is particularly innovative, offering a practical solution for predicting network structures in new conditions. The argumentation is solid, grounded in the speaker’s own research and collaborations, and supported by examples and comparisons with CFD results. The lecture also highlights the limitations of pure ML approaches, such as lack of extrapolation, and proposes a hybrid approach that leverages the strengths of both paradigms.

Scientific Rigor, Source Quality, Title Accuracy

The lecture demonstrates scientific rigor through the use of established principles and references to recent publications, including a paper presented at a symposium and collaborations with MIT. The speaker mentions open-source codes and provides practical resources for further learning. The title accurately reflects the content, which progresses from fundamental principles to advanced ML-enhanced modeling. The lecture is well-structured and technically detailed, though it assumes a high level of prior knowledge. The sources cited are credible, and the speaker is transparent about the limitations and uncertainties of the methods discussed.

The title accurately reflects the content, which covers turbulent combustion modeling from fundamental principles to machine learning-enhanced approaches.

8/10

Lecture by a recognized expert in combustion modeling, presenting advanced research with references to recent publications and open-source tools. The content is technically rigorous and based on established scientific principles, though it is a lecture rather than a peer-reviewed publication.

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

• 00:00Introduction to multi-fidelity frameworks and limitations of reduced-order models.
• 02:00Discussion on co-kriging and the risk of needing many high-fidelity simulations if the low-fidelity manifold is wrong.
• 04:00Explanation of 30-fold statistics and model averaging for robust predictions.
• 06:00Comparison of multi-fidelity digital twin predictions with original CFD for different numbers of high-fidelity simulations.
• 08:00Demonstration that the model cannot extrapolate beyond training data, with an example of NOx prediction from ammonia combustion.
• 10:00Introduction to the COFFEE approach from MIT, using autoencoders and latent space encoding.
• 12:00Mention of open-source tools: OpenMeasure for dimensionality reduction and multi-fidelity ROM.
• 14:00Description of an exercise on dynamic POD for hydrogen flashback prediction.
• 16:00Announcement of a training school on digital twins and a hackathon in Brussels.
• 18:00Introduction to chemical reactor networks (CRNs) as a physics-based approach for digital twins.
• 20:00Explanation of using unsupervised clustering to partition CFD fields into reactor regions.
• 22:00Discussion on the impact of clustering algorithm and variables on the CRN structure.
• 24:00Comparison of CRN predictions with CFD for different hydrogen concentrations in a furnace.
• 26:00Introduction to graph neural networks (GNNs) for predicting CRN topology.
• 28:00Explanation of using GNNs to predict split ratios and network structure for new conditions.
• 30:00Discussion on the training process for GNNs, including masking and feature matrices.
• 32:00Details on predicting topology, split ratios, and node assignments with GNNs.
• 34:00Clarification that GNNs are used only for structure prediction, not for solving physics.
• 36:00Mention of the importance of physics-based solvers for the final CRN simulation.
• 38:00Conclusion and summary of the lecture's key points.

• OpenMeasure — Open-source code for dimensionality reduction, GPR, sparse sensing, and multi-fidelity ROM.
• COFFEE approach (MIT) — Mentioned as a recent approach from Professor Deng's group, using autoencoders and latent space encoding.
• Paper on CRN-GNN (unpublished) — Recent work by the speaker's group on using graph neural networks to predict chemical reactor network structure.

• OpenMeasure — Open-source tool for dimensionality reduction and multi-fidelity modeling, consistent with the lecture's methods.