Physics-informed and data-driven neural networks with dimensional and non-dimensional inputs for single-droplet evaporation: Investigating the role of increasing physical complexity in predictive ability

• Physics-informed neural networks outperform purely data-driven ANNs, especially under complex single-droplet evaporation conditions. • Purely data-driven approaches exhibit notable performance decay when extrapolating beyond training regimes, highlighting the importance of physical constraints. • Integration of Stefan flow and convective effects boosts model fidelity and enables more effective assessment of ANN robustness and reliability. This study extends the investigation of artificial neural network (ANN) models for single droplet evaporation by incorporating Stefan flow and convective effects for more realistic scenarios. The objectives are to assess how these complexities affect ANN accuracy and to see if previous modeling approaches remain effective. Four architectures were compared: conventional no-physics, non-dimensional, physics-informed, and hybrid physics-informed non-dimensional. Data generation covered increasingly complex evaporation, from diffusion-limited to convective mass transfer. Key metrics (R 2 , θ) evaluated interpolation and extrapolation on unseen data. Physics-informed models consistently outperformed others, achieving near-perfect interpolation (R 2 ≈ 1) and robust extrapolation (θ > 90 %). Notably, in the most complex scenario, physics-informed models attained θ = 93.976 %, while no-physics models fell to θ = 26.181 %. Although non-dimensional inputs offered moderate gains, they lagged behind physics-informed models in extrapolation. These findings underscore the value of embedding domain-specific physics to enhance ANN reliability for complex phenomena.