Shape-informed Operator Learning || Separable PINNs|| Nov 7, 2025

Shape-informed Operator Learning || Separable PINNs|| Nov 7, 2025

by CRUNCH Group: Home of Math + Machine Learning + X

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning
🎙 CRUNCH Group: Home of Math + Machine Learning + X 👥 4K 📅 November 7, 2025 ⏱ 121 min 👁 425 📄 expert opinion 🧭 2026-08-15
Available in: English (current) Français

Keywords

operator learningPINNsshape encodingsurrogate modelsscientific machine learning

Summary

This seminar features two talks on advanced machine learning methods for scientific computing. The first talk by Francesco Regazzoni introduces a novel framework for shape-informed operator learning, addressing the challenge of building surrogate models that generalize across varying geometries. The method, called Universal Solution Manifold Networks (USMNs), uses neural fields conditioned on physical parameters and a learned shape code, enabling mesh-free predictions on unseen geometries. Key components include a universal coordinate system based on solving Laplace problems, and automatic shape encoding via auto-decoder neural ODEs or deep SDFs. Results demonstrate accurate predictions for fluid dynamics and solid mechanics problems, with robustness to incomplete geometry representations and topology changes. The second talk by Jaemin Oh presents Separable Physics-Informed Neural Networks (SPINNs), which efficiently solve PDEs on tensor-product domains. Applications include the high-dimensional Boltzmann-BGK model and data assimilation for inferring unknown initial conditions from noisy observations, achieving an 88% reduction in computational time. Both talks highlight the potential of these methods to accelerate simulations and enable many-query tasks.

166 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talks provide substantial value by presenting novel methodologies with clear motivations and rigorous experimental validation. Regazzoni’s argumentation is strong: he systematically identifies bottlenecks in traditional simulation, proposes a comprehensive framework, and supports it with multiple test cases (bifurcation flow, perforated plate, left ventricle mechanics). The use of universal coordinates and learned shape codes is well-justified with theoretical reasoning and empirical comparisons. Oh’s talk is more concise but equally valuable, demonstrating SPINN’s efficiency on challenging problems. The argumentation is solid, though some details are omitted due to time constraints.

98 words

Title / Content Match

The title accurately reflects the content: two talks on operator learning and separable PINNs, presented on Nov 7, 2025.

Quality & Reliability

8/10

The seminar presents two research talks by established academics (Francesco Regazzoni, Politecnico di Milano; Jaemin Oh, Texas A&M University) with peer-reviewed work. The content is technical and detailed, with clear methodology and results. However, as a seminar recording, it lacks formal peer review of the presentation itself, and some claims are not fully verifiable from the video alone.

Key Moments

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

Contribution & Novelties

The seminar presents two significant contributions to scientific machine learning. Regazzoni’s USMN framework introduces a mesh-free, shape-aware operator learning method that can generalize to unseen geometries without retraining, addressing a key limitation of existing approaches. The use of universal coordinates and learned shape codes via auto-decoders is innovative and shows strong empirical performance. Oh’s SPINNs offer an efficient alternative to standard PINNs for tensor-product domains, with demonstrated speedups in high-dimensional and data assimilation problems. These methods have potential to accelerate simulations in engineering and biomedical applications.

Pour aller plus loin :

  • Physics-informed neural networks — Background on PINNs.
  • Operator learning — Overview of operator learning methods.
  • Neural ODEs — Foundational paper on neural ODEs, relevant to shape deformation.
  • DeepSDF — Original paper on deep signed distance functions.

127 words

Radar Profile

The radar profile shows high scores across all dimensions, indicating a technically deep, reliable, and information-rich seminar. The strongest aspects are the technical level and information quality, while the quantity of information is slightly lower due to the seminar format.

Reliability 8/10