A spatio-temporal resolution-adaptive Fourier Neural Operator for phase interface prediction in droplet phase-change

To meet the rapidly growing thermal management demands in modern applications, phase-change cooling technologies with higher heat dissipation capabilities have attracted increasing attention. Accurate prediction of phase-change processes is essential for phase-change cooling technologies such as spray and jet impingement cooling. However, establishing a universal phase-change rate model remains challenging due to the inherent complexity of phase-change mechanisms. Commonly used numerical approaches to model phase-change rate exhibit limited accuracy, and their reliance on empirical parameters requires trial-and-error tuning across different values, leading to considerably high computational costs. To address these limitations, this study proposes a Fourier Neural Operator (FNO) enhanced with physics-informed feature engineering, which enables accurate and efficient prediction of the phase interface during phase-change processes. Droplet impingement on a heated substrate is selected as a representative phase-change process, and a high-fidelity multi-physics dataset is constructed based on experimental data. The FNO is trained on this dataset to predict the temporal evolution of droplet morphology from the multi-physics field, producing physically consistent predictions and yielding superior performance compared to convolutional neural network (CNN)-based baseline models. Moreover, the model demonstrates strong generalization across different spatial and temporal discretization, achieving average accuracy above 95% and inference time within 4 ms on completely unseen spatial resolutions and prediction intervals, highlighting its adaptability to diverse computational fluid dynamics (CFD) mesh configurations and timestep control strategies. The proposed spatio-temporal resolution-adaptive FNO offers a promising alternative for phase-change rate modeling in CFD workflows, paving the way for more accurate and computationally efficient simulations. • A Fourier Neural Operator is developed to predict the phase distribution during the droplet phase-change process, achieving average accuracy above 95% and inference time within 4 ms. • A physics-informed feature engineering method is employed to guide the model training based on physical insights, improving the prediction accuracy by up to 1.41%. • A spatio-temporal resolution-adaptive phase-change rate model is established, enabling adaptation to different spatial and temporal discretization schemes used in CFD simulations.