You are on page 1of 4

Engineering Design

Acceleration through AI
using SMT
[the Surrogate Modeling Toolbox]

Joseph Morlier

ICA UMR CNRS 5312

ATELIER IA POUR LES SCIENCES DE L’INGENIERIE 28-29 JUIN 2022 P1


What is Surrogate Modeling ?
A. Forrester, A. Sobester, and A. Keane. “Engineering Design via Surrogate Modelling: A
Practical Guide”. Coll. John Wiley & Sons (2008).

A surrogate model of a function is an approximation of Expensive


Computer simulation: It’s a supervised learning process in AI.
As the surrogate is less costly to evaluate it can be used as a “fast” code in
a Multidisciplinary Design Optimization loop. [or do Uncertainty
Quantification or do Bayesian Optimization etc…]

(Expensive) simulation MDO


Our approach: is
Collaborative !

y (Fast) Surrogate OpenSource code:


DOE github.com/SMTorg/smt
using SMT y=f(x) Documentation:
length(x)<100
smt.readthedocs.io
x_1 x_2 ATELIER IA POUR LES SCIENCES DE L’INGENIERIE 28-29 JUIN 2022 P2
A Brief Overview of SMT
Bouhlel, M. A., Hwang, J. T., Bartoli, N., Lafage, R., Morlier, J., & Martins, J. R. (2019). A Python surrogate
modeling framework with derivatives. Advances in Engineering Software, 135, 102662.

An open source python toolbox for surrogate models: first version of SMT
in July 2017 1.0 version in April 2021. Novelty : focus on derivatives.
Kriging (aka Gaussian Process in Machine learning), ! = #(%, %', !')
Kriging with partial least square (KPLS) := dimension reduction of input space
%', !' Training data
Gradient-enhanced KPLS & neural networks
%, ! Prediction data

SMT python
package:
Sampling {3}
Engineering
Benchmarks {11}
Models {11}
Examples {3}
Apps {8}

ATELIER IA POUR LES SCIENCES DE L’INGENIERIE 28-29 JUIN 2022 P3


pip install smt
dependencies: numpy, scipy, sklearn, Cython with conda

Practice online:
https://github.com/SMTorg/smt/blob/master/tutorial/README.md

SMT Tutorial (linear, quadratic, gaussian process, ...)

Noisy Gaussian process

Multi-Fidelity Gaussian Process

Bayesian Optimization - Efficient Global Optimization to solve unconstrained Artificial Intelligence


problems for Engineers (AI4E)
Mixed-Integer Gaussian Process and Bayesian Optimization to solve
Tutorials with Jupyter
unconstrained problems with mixed variables (continuous, discrete, categorical) Notebooks {5}

Initial Stiffness=f(lf: length of the fibre,Vc: Percentage of carbon fibre in the mixture)

Composites material prediction (AIAA Aviation 2022):


The aim is to build a surrogate model for mixed variables in a hybrid composite problem using
Gower Distance. The trained surrogate model is able to predict the composites characteristics.
Data from: https://www.sciencedirect.com/science/article/abs/pii/S1359835X11000224

Thanks to Paul Saves & Raul Carreira Rufato

ATELIER IA POUR LES SCIENCES DE L’INGENIERIE 28-29 JUIN 2022 P4

You might also like