Modelling flow in hydrocyclones using physics-informed neural networks
(2025)Department of Automatic Control
- Abstract
- Hydrocyclones are important pieces of equipment used within the mineral processing industry, where they are used for sorting ore particles suspended in water based on their size. The turbulent flow within a hydrocyclone is today usually modelled using demanding computational fluid dynamics (CFD) simulations. In this thesis, we explore the use of physics-informed neural networks (PINNs) as an alternative for modelling hydrocyclone flow. The central idea behind PINNs is modifying the loss function of a feed-forward neural network to include loss terms based on the governing physics of the system, which in this case are the Navier-Stokes differential equations. Using CFD data for comparison, we train and evaluate PINN models both with sparse... (More)
- Hydrocyclones are important pieces of equipment used within the mineral processing industry, where they are used for sorting ore particles suspended in water based on their size. The turbulent flow within a hydrocyclone is today usually modelled using demanding computational fluid dynamics (CFD) simulations. In this thesis, we explore the use of physics-informed neural networks (PINNs) as an alternative for modelling hydrocyclone flow. The central idea behind PINNs is modifying the loss function of a feed-forward neural network to include loss terms based on the governing physics of the system, which in this case are the Navier-Stokes differential equations. Using CFD data for comparison, we train and evaluate PINN models both with sparse training data and completely without training data. Though the simple, data-free model failed to capture the complex physics inside the hydrocyclone, the more complex data-based model showed a lot of potential and managed to somewhat predict the values of the different velocity components and pressure inside of the geometry. However, the complex geometry and turbulent nature of the flow proved to be the main challenges for the model, and a considerable amount of further research within the area is needed before PINNs could possibly be employed to accurately predict hydrocyclone flow. (Less)
Please use this url to cite or link to this publication:
http://lup.lub.lu.se/student-papers/record/9186762
- author
- Cohen Tillberg, Meja and Ingerskog, Ida
- supervisor
- organization
- year
- 2025
- type
- H3 - Professional qualifications (4 Years - )
- subject
- report number
- TFRT-6268
- other publication id
- 0280-5316
- language
- English
- id
- 9186762
- date added to LUP
- 2025-05-09 08:22:50
- date last changed
- 2025-05-09 08:22:50
@misc{9186762,
abstract = {{Hydrocyclones are important pieces of equipment used within the mineral processing industry, where they are used for sorting ore particles suspended in water based on their size. The turbulent flow within a hydrocyclone is today usually modelled using demanding computational fluid dynamics (CFD) simulations. In this thesis, we explore the use of physics-informed neural networks (PINNs) as an alternative for modelling hydrocyclone flow. The central idea behind PINNs is modifying the loss function of a feed-forward neural network to include loss terms based on the governing physics of the system, which in this case are the Navier-Stokes differential equations. Using CFD data for comparison, we train and evaluate PINN models both with sparse training data and completely without training data. Though the simple, data-free model failed to capture the complex physics inside the hydrocyclone, the more complex data-based model showed a lot of potential and managed to somewhat predict the values of the different velocity components and pressure inside of the geometry. However, the complex geometry and turbulent nature of the flow proved to be the main challenges for the model, and a considerable amount of further research within the area is needed before PINNs could possibly be employed to accurately predict hydrocyclone flow.}},
author = {{Cohen Tillberg, Meja and Ingerskog, Ida}},
language = {{eng}},
note = {{Student Paper}},
title = {{Modelling flow in hydrocyclones using physics-informed neural networks}},
year = {{2025}},
}