Explainable machine learning maps the efficiency-NOx trade-off of a hydrogen dual-fuel engine fuelled with a diesel–Mahua methyl ester-pine oil blend

Abstract Hydrogen-enriched dual-fuel combustion offers a route to decarbonising compression-ignition (CI) engines, but its interaction with oxygenated ternary fuels remains poorly quantified. This study evaluates a Span 80-stabilised diesel-Mahua methyl ester (MME)-pine oil (PO) blend on a Kirloskar TV1 CI engine (5.2 kW, 1500 rpm, compression ratio 17.5:1). Twenty operating points for D70 + MME15 + PO15 were measured across five loads (20–100%) and four hydrogen rates (0, 5, 10 and 15 L min −1 , equivalent to a 0–15.4% hydrogen energy share at full load); the remaining sixty points of the design (D100, D80 + MME10 + PO10, D60 + MME20 + PO20) were generated with a calibrated zero-dimensional thermodynamic model and are identified as model-predicted throughout. Raising hydrogen from 0 to 15 L min −1 at full load increased brake thermal efficiency (BTE) by 5.6–8.0% and reduced CO by 59.1–63.0%, HC by 59.3–61.5% and smoke by 20.7–26.3%, while NOx rose by 15.0–21.4%; the highest BTE observed was 31.74%. A Kolmogorov-Arnold network (KAN) outperformed XGBoost and an artificial neural network under five-fold cross-validation (mean R 2 = 0.9913). SHAP attribution identified engine load as the dominant driver of BTE (65.7%) and NOx (69.6%), and hydrogen induction as the dominant driver of CO (75.4%) and HC (54.7%), while conformal prediction supplied calibrated 95% intervals. NSGA-II optimisation of the KAN surrogate returned a 118-point Pareto front: an efficiency-priority solution retained 99.4% of the maximum BTE while lowering NOx by 7.4% and smoke by 10.4%, and a balanced solution traded 11.2% of BTE for a 40.9% NOx reduction. All optimisation results are surrogate predictions reported with calibrated 95% intervals and await experimental confirmation.