Description du projet
Apprentissage automatique conscient de la physique pour la mécanique des fluides
La capacité de la modélisation de la mécanique des fluides à prédire l’évolution d’un écoulement est rendue possible à la fois par des principes physiques et des approches empiriques. D’une part, les principes physiques (par exemple les lois de conservation) sont extrapolatifs – ils fournissent des prédictions sur des phénomènes qui n’ont pas été observés, d’autre part, la modélisation empirique fournit des fonctions de corrélation au sein des données. L’intelligence artificielle et l’apprentissage automatique sont d’excellents outils de modélisation empirique. Le projet PhyCo, financé par l’UE, combinera les principes physiques et la modélisation empirique en une approche unifiée: des méthodes d’optimisation multiphysiques basées sur des données limitées par la physique. Les solutions d’apprentissage automatique n’enfreindront pas les contraintes physiques. Le cadre de calcul sera appliqué pour reconstruire des images d’écoulement à haute résolution à partir de données à basse résolution, pour minimiser les émissions des moteurs d’avion avec des écoulements réactifs à base d’hydrogène et pour maximiser la récupération d’énergie sans émission à partir des oscillations de la structure des fluides.
Objectif
About a hundred trillion bytes of data has been created in the world while reading this sentence. Central to big data is machine learning, which is an automated way of transforming information into empirical knowledge. Machine learning techniques have been applied to some fluid mechanics problems with success, but there are still three big open questions: Do machine learning algorithms scale to engineering configurations? (Are they robust?); Can we gain physical insight into the solutions? (Are they interpretable?); Can we extrapolate knowledge to other configurations, such as multi-physics problems? (Are they generalizable?).
Fluid mechanics modelling has been historically enabled by both empirical approaches and physical principles. Machine learning models may not be interpretable and robust, but they excel at empirical modelling. On the other hand, physical principles are governed by equations that do not adaptively change, but they are interpretable and robust. This project will combine physical principles and empirical modelling into a unified approach: physics-constrained adaptive learning for multi-physics optimization of unsteady, unpredictable and uncertain flows. The learned solutions will not violate physical constraints.
The technical objectives are to combine physical principles with machine learning; design adaptive multi-physics models by on-the-fly data assimilation; optimize turbulent flows; quantify the uncertainty; and develop a code that wraps around existing simulation software and experiments. This framework will be applied to maximize energy harvesting from aeroelastic systems to produce clean energy; optimize stable aeroengines with low emissions; and reconstruct high-resolution flow fields from low-resolution experimental measurements. We will rigorously interlace chaos theory, Bayesian inference and artificial intelligence. This project will benefit industries that work with multi-physics flows and artificial intelligence companies.
Champ scientifique
- natural sciencescomputer and information sciencesdata sciencebig data
- natural sciencesmathematicsapplied mathematicsstatistics and probabilitybayesian statistics
- natural sciencescomputer and information sciencescomputational sciencemultiphysics
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
- natural sciencescomputer and information sciencessoftwaresoftware applicationssimulation software
Mots‑clés
- adjoint equations in stability
- least square shadowing
- computational fluid dynamics
- reacting flows
- machine learning in fluid mechanics
- virtualization of flows
- gas turbines
- reduced-order models of fluids
- thermoacoustic instabilities
- aeroelastic instabilities
- combustion instabilities
- acoustic-flow interaction
- proper orthogonal decomposition
- dynamic mode decomposition
- dimensionality reduction
- flow state estimator
- uncertainty quantification in fluid dynamics
- turbule
Programme(s)
Thème(s)
Régime de financement
ERC-STG - Starting GrantInstitution d’accueil
10129 Torino
Italie