Predicting Spray Penetration in High-Pressure Combustion Using Machine Learning Models Tuned on ECN Data

Machine learning (ML) models provide fast, efficient predictors of key spray observables that govern mixture preparation in direct fuel-injection systems, reducing reliance on costly optical diagnostics and high-fidelity simulations. We develop and benchmark supervised ML models to predict steady liquid penetration length (l_l) and transient vapor penetration length (l_v) using standardized ECN Spray A/C/D data curated into N = 333 observations with 19 physical features spanning ambient conditions, injector geometry, and fuel descriptors. We compare several classical regressors—linear regression, support vector regression (SVR), gradient boosting (GB), random forest (RF), and extreme gradient boosting (XGB)—with ML-based neural models: multilayer perceptron (MLP), convolutional neural network (CNN), temporal convolutional network (TCN), and recurrent neural network (RNN). For steady l_l, tree-based ensembles (especially RF and GB) are the most accurate and stable, yielding mean absolute error (MAE) about 2.1–2.2 mm with coefficient of determination R^2 ≈ 0.94. For transient l_v, the TCN gives the best overall accuracy (MAE ≈ 3.0 mm; R^2 ≈ 0.68). Feature-importance and attribution analyses identify ambient density, nozzle diameter, and injection pressure as the dominant drivers, consistent with atomization and vaporization physics. Runtime benchmarks show higher training cost for deep models than for classical regressors, while inference time is comparable. We further assess generality by applying models trained on diesel data to kerosene surrogates (RP-1/RP-2). Tree-based regressors predict steady l_l within ~3% of reference values, whereas for l_v the TCN achieves ~4–5%. Overall, tree-based regressors are adequate for steady l_l, while TCN-type architectures are more accurate for transient l_v.