CrayStack ensemble machine learning for predicting engine performance and emissions of multiple feedstock biodiesel blends with CaO-Al2O3 nanoparticles

Abstract

Abstract The global transition toward sustainable energy has positioned multi-oil feedstocks (Garcinia gummi-gutta and Garcinia indica) to increase feedstock availability, which is vital to biodiesel production, thereby reducing fossil fuel dependencies and limiting harmful emissions. The incorporation of multifunctional CaO-Al 2 O 3 nanoparticles into biodiesel further enhances biodiesel combustion characteristics. However, predicting an engine performance (brake thermal efficiency: BTE, brake specific fuel consumption: BSFC) and emissions (carbon monoxide: CO, unburnt hydrocarbon: UHC, and nitrogen oxide: NOx) remains a critical bottleneck for currently practiced traditional and resource-intensive experimental methods due to complex interplay of fuel properties (biodiesel blends: B10, B20, B30, and nanoparticles concentration possessing an average particle size of 75.4 nm) and engine variables (engine load and compression ratio). To bridge this gap, this study used a CrayStack machine learning ensemble, which integrates three diverse base learners—Gaussian Process Regression (GPR), Least Squares Boosting (LSBoost), and Support Vector Regression (SVR)—with a meta-learner whose ensemble weights are dynamically optimized via the cutting-edge Crayfish Optimization Algorithm (COA). This nature-inspired optimization method uniquely captures complex, multi-response nonlinearities using highly constrained experimental datasets (25 training, 10 testing), a task that conventional models fail to perform. In the CrayStack ensemble framework, COA evaluates the contributions of three base learners, with GPR having the highest contribution (> 50%) for BTE, BSFC, CO, and NOx, and LSBoost accounting for 52.3% for UHC. The Pairwise Wilcoxon single-ranked analysis confirmed that the CrayStack ensemble consistently outperformed SVR for all five responses. The CrayStack ensemble demonstrated superior predictive performance, with consistent results across all responses and the lowest composite Mean Absolute Percentage Error (MAPE) of 4.38%, outperforming GPR (4.93%), SVR (6.01%), and LSBoost (7.70%). Integrating CrayStack Ensemble with Monte Carlo simulations (MCS) yields probabilistic predictions that consistently outperform three base learner models, with the lowest error (RMSE of 0.35 and standard deviation of 0.02), ensuring robust and reliable predictions across all responses. CrayStack ensemble requires an average computational time of 3.736s, demonstrating its computational efficiency, robustness, and reliability for accurate engine performance and emission prediction applications. Thereby, CrayStack drastically eliminated experimental overhead (cost, power, materials) and provided a transformative, scalable tool to accelerate the commercial deployment of nanoparticle-enhanced sustainable biofuel blends.