Fiche de QU Jingang

Titre : Acceleration of Numerical Simulations with Deep Learning : Application to thermodynamic equilibrium calculations
Date : 29 juin 2023
Heure: 09:00
Résumé : The primary objective of this thesis is to accelerate thermodynamic equilibrium calculations using machine learning and deep learning techniques. Thermodynamic equilibrium calculations are used to identify the phases of mixtures and their composition at equilibrium and play a pivotal role in many fields, such as chemical engineering and petroleum industry. To achieve this objective, we leverage PyTorch to rewrite and vectorize algorithms for high-performance computing (HPC), while also using neural networks to replace repetitive and time-consuming subroutines. Additionally, this thesis focuses on domain generalization (DG) in image classification. DG involves training models on known domains that can effectively generalize to unseen domains. However, current DG methods often require domain labels and lack interpretability. Therefore, we aim to develop a novel DG method that does not require domain labels and is more interpretable.
Lieu : 4 Pl. Jussieu, 75005 Paris, Sorbonne University Pierre and Marie Curie Campus