Improving aerosol and spray process computation fluid dynamics models with machine learning approaches

Computational fluid dynamic (CFD) models of aerosol and sprays are used across a wide range of fields; from understanding dispersion of exhaled aerosols, to designing and optimising spray drying process for production of food, pharmaceuticals and consumer goods. These CFD models often use particle scale models of drying kinetics to capture the dynamics and solidification behaviour as the droplets interact with the air flow.
The drying behaviour of single droplets can be complex due to nature of the phase changes occurring and interaction between the drying behaviour and the solid structures formed on drying. Models which capture the complexity of the drying behaviour have been developed for single particles, however integrating these models into a CFD model with multiple particles present is typically limited by computational time. The drying models used in CFD models are therefore simple approximations which do not capture the complex physics involved in drying and solid formation.
There is therefore a need to develop models which capture the more complex physics but that run faster and can therefore be used in CFD models. One recent innovation, that offers some potential to be able to do this, is the development of machine learning approaches in fluid mechanics. Amongst the numerous methods available are Physics Informed Neural Networks4 (PINNs), these are neural network models which preserve the physical constraints present in first principle models, but can be executed far more rapidly than direct solution of the differential equations involved in the model. This project will explore the potential of these methods. Initially PINN models of single particles will be developed and their accuracy and speed tested These models will then be incorporated into a CFD model of a spray drying tower and their performance evaluated.