Improving simulations of large-scale dense particle laden flows with machine learning: a genetic programming approach

• Applicants: Professorin Dr.-Ing. Sanaz Mostaghim; Professor Dr. Berend van Wachem
• Subject Area: Fluid Mechanics ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing ; Mechanical Process Engineering
• Term: since 2021
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 466092867

Particle-laden flows are encountered in many natural and industrial processes, such as, for instance, the flow of red and white blood cells in plasma, or the fluidization of biomass particles in furnaces. Over the last 40 years, scientists have used Euler-Lagrange (EL) simulations as a way to predict the behaviour of such flows. However, EL simulations rely heavily on closure models to describe the sub-filter stresses (i.e. fluid turbulence) and the interaction between the fluid and the particles. These closure models arise from complex physical phenomena occurring at the small scales: how small fluid eddies interact, and how the fluid flow interacts with the surfaces of the particles. Current models for this are largely empirical, rudimentary, and precisely determining values for these would be extremely expensive, as it would involve running highly resolved simulations for each case. This is a proposal to deliver novel models for the sub-filter stresses and the particle-fluid interactions for predicting particle-laden flows at the process level, given the properties of the flow around the particle and of the surrounding particles, using a supervised machine learning approach: genetic programming (GP). GP is highly suitable, as its results in verifiable expressions for the aforementioned closures. In the first funding period, we have been able to deliver a new particle-fluid interaction model for the Stokes regime, which is very accurate. In this funding period, we will extend the fluid-particle interaction model to include uncertainty quantification, and develop an expression for the unclosed sub-filter stresses. The developed expressions will be validated by analytical solutions and highly resolved simulations, and will enable accurate, large-scale simulations of dense particle-laden flows at the process level, while only requiring a fraction of the cost of fully resolved simulations

• DFG Programme: Priority Programmes
• Subproject of: SPP 2331: Machine Learning in Chemical Engineering. Knowledge Meets Data: Interpretability, Extrapolation, Reliability, Trust
• International Connection: USA
• Cooperation Partner: Professor Dr. Fabien Evrard