New Oxford AI tool predicts droplet behaviour from spray images alone

Researchers at the University of Oxford have developed a new artificial intelligence framework that can predict the microscopic behaviour of liquid sprays from images alone — reducing calculations that would normally take hundreds of simulation hours to just seconds.

The study, led by Professor Konstantina Vogiatzaki, focuses on “jet-in-crossflow” sprays, where a liquid jet is injected into a fast-moving gas stream. These sprays are widely used in technologies ranging from aircraft engines and industrial cooling systems to fuel injection and clean-energy applications.

One of the biggest challenges in understanding sprays is predicting the sizes of the droplets they produce. Droplet size strongly affects how efficiently fuels burn, how cooling systems perform, and how pollutants spread. However, accurately modelling these microscopic droplets usually requires extremely computationally intensive simulations.

The Oxford team tackled this challenge by combining high-fidelity fluid simulations with deep learning. Their new two-stage AI system learns how to connect large-scale spray images - the visible “macroscopic” behaviour of the jet - with the underlying microscopic droplet-size distributions.

The researchers trained the system using thousands of synthetic images generated from advanced large eddy simulations (LES), a state-of-the-art computational fluid dynamics technique. The AI first analyses spray images to identify key flow conditions, before predicting the full distribution of droplet sizes and associated uncertainties.

Crucially, the framework can produce results in seconds while maintaining high accuracy, reducing computational cost by more than 99.9% compared with conventional simulation methods.

Professor Vogiatzaki says, “What is exciting about this work is that we demonstrated, for the first time, that macroscopic synthetic image data can be used to accurately predict microscopic droplet statistics almost instantly. Traditionally, obtaining this type of information would require very expensive numerical simulations running for hundreds of simulations hours.”

The study also showed that the AI could successfully reproduce complex spray behaviours, including multimodal droplet distributions, without assuming a predefined mathematical form. This means the framework can adapt to highly complex flows that are difficult to model using traditional approaches.

“This work showcases beautifully how AI can augment and accelerate excellent science and engineering. Obtaining higher-accuracy results at a tiny fraction of the compute cost is a real achievement.”

Professor Stephen Roberts

The researchers believe the method could eventually help engineers design more efficient combustion systems, improve industrial spraying technologies, and accelerate the development of lower-emission energy systems.

Although the current study was trained entirely on simulation-generated images, the framework was specifically designed with future experimental use in mind. The team hopes the approach could eventually be applied to high-speed camera footage from real-world spray systems, enabling rapid diagnostics and real-time analysis.

The research was carried out by Professor Konstantina Vogiatzaki, Professor Stephen Roberts, Postdoctoral Research Assistant Dr Giovanni Tretola, and Luke Vernon at the Department of Engineering Science.

Dr Tretola played a key role in developing the research, bringing specialist expertise at the interface of multiphase-flow physics, numerical simulation, and machine learning. Reflecting on the publication, he says, “I am very pleased that this work has been published, as it brings together two areas I care deeply about: multiphase-flow physics and machine learning.”

The initial work for the paper was carried out by Luke Vernon, who recently graduated from the Oxford MEng undergraduate course, as part of his 4th Year Project. Luke made an important contribution to the development of the framework, demonstrating how undergraduate research can feed directly into high-quality research outputs. Commenting on the experience, Luke says, “Working on this project as part of my 4YP alongside such a great team, was an incredibly rewarding experience. Having the opportunity to present the research at a conference and see the work published also speaks to the real-world impact and value of the research.”

“This is a wonderful example of how undergraduate research can develop into a publication in a leading journal.”

Professor Konstantina Vogiatzaki

The paper, “A Computational Framework for Droplet Size Distribution in Liquid Jets into Crossflow via Deep Learning and Image-Based Analysis”, has been accepted into the Journal of Computational Physics, a leading journal in computational research. Full paper.