Machine learning computational study on solubility estimation of hydrogen in diesel
This study develops and evaluates three regression approaches for predicting hydrogen (H 2 ) solubility in catalytic diesel from temperature and pressure: Decision Tree (DT), Gamma Regression (GR), and Multilayer Perceptron (MLP). The contribution is a leakage-controlled assessment of these models on a small, structured experimental dataset, together with Harmony Search (HS) hyperparameter optimization and comparison against conventional regression and machine-learning benchmarks. The 30 observations cover 353.2–453.2 K and 1.15–5.86 MPa. Generalization was evaluated by nested leave-one-temperature-level-out cross-validation: each temperature level served once as an outer test group, while HS operated exclusively within the remaining training groups. From pooled out-of-fold predictions, DT, GR, and MLP achieved R 2 values of 0.732, 0.959, and 0.971; RMSE values of 0.752, 0.294, and 0.245; MAE values of 0.593, 0.254, and 0.201; and MAPE values of 21.65%, 9.99%, and 6.52%, respectively. The MLP therefore provided the best overall predictive performance, although the ablation analysis showed that HS materially improved GR but did not uniformly improve DT or MLP under unseen-temperature testing. Pressure was the dominant sensitivity for all three models. The validated applicability domain is restricted to catalytic diesel within the experimental temperature-pressure range, where the models may support preliminary hydroprocessing calculations but do not replace equilibrium measurements outside that domain.
Authors
- Ahmed Bakr Khoshaim (ORCID: https://orcid.org/0000-0002-1409-8120)
- Almuhannad Sulaiman Alorfi (ORCID: https://orcid.org/0009-0009-3453-4127)
Institutions
- King Abdulaziz University (SA)
Publication Details
- Journal
- Frontiers in Chemistry
- Published
- 2026-11-20
- DOI
- https://doi.org/10.3389/fchem.2026.1957487
- Primary Topic
- Phase Equilibria and Thermodynamics
- Type
- article
- Field-Weighted Citation Impact
- 0.00