Title:	Machine Learning Methods for Non-Linear Model Reduction of Turbulent Flows and its Prediction
DNr:	LiU-gpu-2026-1
Project Type:	LiU
Principal Investigator:	Aditya Prakash Ghosh <aditya.prakash.ghosh@liu.se>
Affiliation:	Linköpings universitet
Duration:	2026-02-20 – 2026-07-01
Classification:	20306
Homepage:	https://saeedsalehi.com/opportunities.html
Keywords:

Abstract

This thesis investigates Cuda accelerated non linear model reduction and forecasting for turbulent fluid flow dynamics using data driven latent variable modelling. The primary focus is on 2D and 3D vortex rope simulation datasets generated from scale resolving CFD (DDES), where the resulting transient fields are high dimensional, multi scale, and computationally expensive to analyze and reuse for prediction. To address these challenges, the thesis develops a workflow that compresses full order flow fields into compact latent representations while retaining the essential non linear dynamics needed for accurate reconstruction and time evolution. The core model reduction component is based on beta variational autoencoders (VAEs) trained on time resolved flow snapshots on both 2D and 3D structured mesh. The VAE learns a structured latent space that enables faithful reconstruction of key flow features (including coherent vortical structures) and supports controllable compression ratios across mesh resolutions and dimensionality. Building on this reduced representation, the thesis then targets transient prediction directly in latent space to enable fast rollouts of vortex rope dynamics without repeatedly running full CFD. Two complementary forecasting families are explored: (i) recurrent sequence models (LSTM based predictors) for deterministic or probabilistic latent trajectories, and (ii) diffusion based generative temporal models for uncertainty aware multi-step forecasting and robust long horizon rollouts.