Abnormal convergence issue in DeepONet || ML in modeling turbulent mixing || Sep 5, 2025
Keywords
Summary
127 words
Critical Evaluation
Value of the Information & Strength of the Argument
The first talk provides a clear motivation for the proposed method, grounded in the universal approximation theorem for operators and the analogy to basis expansion methods in physics. The argumentation is logical: identify the issue (linear dependence of learned basis), propose a fix (orthogonalization), and demonstrate improvement on multiple examples. The speaker acknowledges limitations and open questions, which adds credibility. The second talk is less detailed in the transcript, but the topic is relevant and likely presents valuable insights into ML for turbulent mixing. Overall, the seminar offers substantial technical content and critical discussion.
103 words
Title / Content Match
The title accurately reflects the content: two talks on machine learning in scientific computing, focusing on DeepONet convergence and turbulent mixing.
Quality & Reliability
7/10
The seminar presents original research on improving DeepONet convergence, with technical depth and critical discussion. However, the video is a recording of a seminar with limited production quality, and the claims are not peer-reviewed in this context.
Key Moments
Markers derived by PSI from the transcript: the creator did not define chapters.
- Introduction and first talk begins
- Jie Zhao discusses motivation from physics and operator learning
- Explanation of DeepONet architecture and universal approximation
- Demonstration of abnormal convergence issues with examples
- Proposed modification: truncation and QR decomposition of trunk net
- Results showing improved convergence and accuracy
- Q&A session begins, discussion on computational cost and relation to other methods
- Second talk by Sébastien Thévenin on ML for turbulent mixing
- Discussion and further questions
Contribution & Novelties
The seminar presents a novel modification to DeepONet to address convergence issues by enforcing orthogonality of the learned basis functions via QR decomposition. This is a practical contribution that could improve the reliability of neural operators. The discussion also highlights open questions and comparisons with other methods.
Pour aller plus loin :
- DeepONet: Learning nonlinear operators — Original paper introducing DeepONet.
- Universal approximation theorem for operators — Theoretical foundation.
- Proper orthogonal decomposition — Related basis reduction technique.
77 words
Radar Profile
The radar profile shows high scores in technical level and information quantity, reflecting the seminar's depth. Quality and reliability are moderate due to lack of explicit citations and informal setting. The overall balance indicates a valuable but not fully polished scientific communication.