A discussion of the current and potential uses of Gaussian Processes in mechanics
Résumé
Data-driven approaches to modeling and design in mechanics often assume, when relying on neural networks, that a large amount of data is available. However, this is not always the case as the experiments or the physical simulations may be time or money consuming. Data then tends to be small rather than big.
Gaussian processes (GPs) are a versatile statistical model that is relevant for small data thanks to an intrinsically interpolating structure. Modeling with a GP through kernel design allows both flexibility and some explainability.
In this talk, we will show how kernel design can be useful in mechanics in three challenging situations: costly simulations and
i) high dimension (shapes optimization) ii) discrete inputs (e.g., materials taken in a database) iii) intrinsic randomness (permeability fields).
Origine | Fichiers produits par l'(les) auteur(s) |
---|