Physics-informed vs. data-driven neural networks for n-heptane droplet evaporation

The performance of three artificial neural network (ANN) models including a pure data-driven, a feature-engineered, and a hybrid physics-informed model for predicting n-heptane droplet evaporation across increasingly complex conditions, from pure diffusion to cases incorporating Stefan flow and convective effects was investigated. The training dataset was selected to span effective ranges of Stefan flow and convective effects. Model performance was assessed under completely unseen conditions, representing an extrapolation regime. Across all cases, the physics-informed ANNs consistently outperformed the other models, achieving near-perfect interpolation (R² > 0.999) and exceptional extrapolation performance, with 100.00% of predictions falling within a ±20.00% error band (θ), even in the most complex convective case. In contrast, the purely data-driven model showed a substantial degradation in extrapolation capability for the convective case (θ = 47.6 ± 9.6%). Feature-engineered models yielded moderate improvements for diffusion- and Stefan-flow-dominated cases (θ = 86.7 ± 23.1% and 61.5 ± 10.2%, respectively). For the convective case, however, a high extrapolation accuracy (θ = 97.9 ± 2.5%) was achieved owing to the inclusion of velocity-dependent non-dimensional input groups that effectively captured convective transport effects. These results underscore the advantage of embedding domain-specific physics and carefully selected input features for accurate prediction of complex transport phenomena. • Physics-informed models achieved 100% extrapolation accuracy up to 130°C beyond training. • Data-driven models failed in extrapolation (θ = 28.5-47.6%) exhibiting massive error growth. • Dimensionless feature engineering recovered 61.5-97.9% accuracy without full physics. • Physics-informed models kept low RMSPE across temperatures; data-driven errors increased.