12–14 Oct 2021
on-line
Europe/Amsterdam timezone

TMM Reduction using Machine Learning Techniques

13 Oct 2021, 12:00
30m
on-line

on-line

thermal analysis and software tools Thermal Analysis

Speaker

Vincent Vadez (DOREA)

Description

A recurrent industrial problematic lies in integrating sub-models into full and complex models, such as adding a spacecraft model into a launcher model. Thermal nodes clustering is currently proceeded manually by thermal knowledge from the technical expert, taking a substantial amount of time.
One of the milestones of the PhD funded by Dorea about Model Reduction for Space Thermal Simulation includes the study of state-of-the-art machine learning and deep learning algorithms, notably clustering algorithms (e.g.: k-means, spectral clustering, mean-shift). Deep learning being a subset of machine learning, the main difference between them is how feature extraction is handled. For usual machine learning algorithms, the user manually chooses relevant features, check whether the output is as required, and adjust the algorithm if this is not the case, which seemed relevant for this study. For deep learning algorithms, however, features are extracted automatically, and the algorithm learns from its own errors.
The idea detailed here consists in elaborating a proof of concept about automatic clustering of thermal nodes into average nodes (according to TMRT terminology) via machine learning. Depending on the AI method chosen by the user, the number of clusters (consisting of a grouping of thermal nodes), can either be the input or the output of the clustering techniques.
The first naive approach considers four parameters for each point associated to a thermal node: the coordinates of its centre of mass (x,y,z) and its associated temperature T for a given configuration at a fixed time step. Then clustering machine learning algorithms from the scikit-learn library are tested on these input data, resulting in different clustering configurations. The results induced by the reduced models are compared and discussed with the original calculations case.
If the results of the concept are encouraging, a tool that considers multiple configurations over a complete orbit could be developed.

Primary authors

Presentation materials