Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning large scale industrial physics simulations

Fabien Casenave

TL;DR

The paper tackles the challenge of expensive, high-dimensional physics simulations in industrial design by presenting two complementary strategies: (i) physical reduced-order modeling (ROM) that uses snapshot POD and hyperreduction to accelerate nonlinear structural mechanics and transient heat analyses, and (ii) Mesh Morphing Gaussian Process (MMGP), which uses shape embedding, mesh morphing, FE interpolation, PCA, and Gaussian process regression to learn solutions with non-parameterized geometrical variability. It demonstrates substantial practical benefits, including orders-of-magnitude speed-ups and uncertainty-aware predictions, and validates MMGP on industrial-style problems such as AirfRANS, where it achieved top performance in a competition. The work further contributes to openness by releasing open-source libraries (PLAID, MMGP, Mordicus, genericROM, Lagun) and non-confidential datasets, fostering reproducibility and community engagement. Overall, the approaches provide scalable, physics-informed surrogates that enable rapid design optimization and reliability assessment for large-scale industrial simulations.

Abstract

In an industrial group like Safran, numerical simulations of physical phenomena are integral to most design processes. At Safran's corporate research center, we enhance these processes by developing fast and reliable surrogate models for various physics. We focus here on two technologies developed in recent years. The first is a physical reduced-order modeling method for non-linear structural mechanics and thermal analysis, used for calculating the lifespan of high-pressure turbine blades and performing heat analysis of high-pressure compressors. The second technology involves learning physics simulations with non-parameterized geometrical variability using classical machine learning tools, such as Gaussian process regression. Finally, we present our contributions to the open-source and open-data community.

Learning large scale industrial physics simulations

TL;DR

The paper tackles the challenge of expensive, high-dimensional physics simulations in industrial design by presenting two complementary strategies: (i) physical reduced-order modeling (ROM) that uses snapshot POD and hyperreduction to accelerate nonlinear structural mechanics and transient heat analyses, and (ii) Mesh Morphing Gaussian Process (MMGP), which uses shape embedding, mesh morphing, FE interpolation, PCA, and Gaussian process regression to learn solutions with non-parameterized geometrical variability. It demonstrates substantial practical benefits, including orders-of-magnitude speed-ups and uncertainty-aware predictions, and validates MMGP on industrial-style problems such as AirfRANS, where it achieved top performance in a competition. The work further contributes to openness by releasing open-source libraries (PLAID, MMGP, Mordicus, genericROM, Lagun) and non-confidential datasets, fostering reproducibility and community engagement. Overall, the approaches provide scalable, physics-informed surrogates that enable rapid design optimization and reliability assessment for large-scale industrial simulations.

Abstract

In an industrial group like Safran, numerical simulations of physical phenomena are integral to most design processes. At Safran's corporate research center, we enhance these processes by developing fast and reliable surrogate models for various physics. We focus here on two technologies developed in recent years. The first is a physical reduced-order modeling method for non-linear structural mechanics and thermal analysis, used for calculating the lifespan of high-pressure turbine blades and performing heat analysis of high-pressure compressors. The second technology involves learning physics simulations with non-parameterized geometrical variability using classical machine learning tools, such as Gaussian process regression. Finally, we present our contributions to the open-source and open-data community.

Paper Structure

This paper contains 12 sections, 10 figures.

Table of Contents

  1. Introduction
  2. Physical Reduced-Order Modeling
  3. Linear data compression
  4. Piecewise linear data compression
  5. Limitations
  6. Learning physical problems with non-parameterized geometrical variability
  7. Mesh Morphing Gaussian Process
  8. ML4PhySim competition
  9. Outlooks
  10. Open-source and open-data
  11. Open-source codes
  12. Open-data datasets

Figures (10)

  • Figure 1: Optimization of the shape of an aircraft wing.
  • Figure 2: Piecewise linear data compression
  • Figure 3: Local ROM recommendation.
  • Figure 4: Some output fields for the 2D advection problem daniel2022physics.
  • Figure 5: Number of POD modes for each local basis with respect to the number of clusters, for various accuracy criterions of the POD truncature, applied to the 2D advection problem daniel2022physics.
  • ...and 5 more figures