A comparative study of dimensional and non-dimensional inputs in physics-informed and data-driven neural networks for single-droplet evaporation

• Physics-informed neural network effectively predicts single-droplet evaporation. • Non-dimensional inputs improve extrapolation to unseen conditions. • Results highlight the role of domain knowledge in robust scientific ML models. • Study provides a framework for applying physics-informed ML to heat and mass transfer problems. This study explores the potential of incorporating physically meaningful non-dimensional inputs and physical constraints into data-driven models to improve prediction efficiency. Using single liquid droplet evaporation as a case study, five artificial neural network models were developed: a no-physics model, a physics-guided model, a physics-informed model with governing equations, a non-dimensional model using non-dimensional inputs, and a physics-informed non-dimensional model combining both non-dimensional inputs and physical constraints. The physics-informed model achieved 86.61 % of predictions within the ±20 % error band during extrapolation, compared to 41.07 % for the no-physics model. Non-dimensional inputs significantly improved extrapolation capability, with 61.61 % accuracy. Although the physics-informed non-dimensional model was less accurate, it demonstrated greater consistency, with a lower standard deviation (10.19) compared to the physics-informed model (48.79). These results emphasize the importance of data representation and domain knowledge in developing robust and generalizable machine learning models for scientific applications