Generative Models for the Deformation of Industrial Shapes with Linear Geometric Constraints: model order and parameter space reductions

by   Guglielmo Padula, et al.

Real-world applications of computational fluid dynamics often involve the evaluation of quantities of interest for several distinct geometries that define the computational domain or are embedded inside it. For example, design optimization studies require the realization of response surfaces from the parameters that determine the geometrical deformations to relevant outputs to be optimized. In this context, a crucial aspect to be addressed are the limited resources at disposal to computationally generate different geometries or to physically obtain them from direct measurements. This is the case for patient-specific biomedical applications for example. When additional linear geometrical constraints need to be imposed, the computational costs increase substantially. Such constraints include total volume conservation, barycenter location and fixed moments of inertia. We develop a new paradigm that employs generative models from machine learning to efficiently sample new geometries with linear constraints. A consequence of our approach is the reduction of the parameter space from the original geometrical parametrization to a low-dimensional latent space of the generative models. Crucial is the assessment of the quality of the distribution of the constrained geometries obtained with respect to physical and geometrical quantities of interest. Non-intrusive model order reduction is enhanced since smaller parametric spaces are considered. We test our methodology on two academic test cases: a mixed Poisson problem on the 3d Stanford bunny with fixed barycenter deformations and the multiphase turbulent incompressible Navier-Stokes equations for the Duisburg test case with fixed volume deformations of the naval hull.


page 8

page 9

page 11

page 13

page 18

page 19

page 20

page 25


On Linear Interpolation in the Latent Space of Deep Generative Models

The underlying geometrical structure of the latent space in deep generat...

Analysis-aware defeaturing of complex geometries

Local modifications of a computational domain are often performed in ord...

Deep Markov Chain Monte Carlo

We propose a new computationally efficient sampling scheme for Bayesian ...

A two stages Deep Learning Architecture for Model Reduction of Parametric Time-Dependent Problems

Parametric time-dependent systems are of a crucial importance in modelin...

Sampling-free parametric model reduction for structured systems

We consider the reduction of parametric families of linear dynamical sys...

Universal Solution Manifold Networks (USM-Nets): non-intrusive mesh-free surrogate models for problems in variable domains

We introduce Universal Solution Manifold Network (USM-Net), a novel surr...

Please sign up or login with your details

Forgot password? Click here to reset