2026 Guida Spray Injector Neural Surrogate

Spray atomisation is used in a variety of applications that rely on its ability to instantaneously create an extremely large surface area between the liquid and gas phases. However, predictive analysis of spray behaviour and estimates of surface area are extremely complex. Experimental activities are constrained by diagnostics that cannot access all spray regions, while numerical methods are computationally expensive, particularly as finer structures form. Data-driven methods can address this issue by learning how interfaces are generated and evolve in space, enabling users to replace complex, often slow CFD simulations with fast inference, thereby quickly exploring design space, ranking conditions, and ultimately controlling spray atomisation. In this work, we propose an architecture that learns the most relevant parameter in sprays: their liquid-gas surface evolution. The principal method we use is a boundary-conditioned Fourier Neural Operator that learns the evolution of the Signed Distance Function from the gas-liquid interface, trained on a dataset spanning several atomisation regimes. Our evaluation shows that the proposed SDF-FNO model retains better fidelity than traditional methods while preserving the ordering of operating points.