Rapid Fuel Distribution Prediction in RCCI Marine Engines Using CFD-Informed, Active Learning Framework

Accurate prediction of in-cylinder fuel distribution (FD) is fundamental to reduced-order combustion modeling and emissions prediction, yet remains computationally prohibitive with high-fidelity CFD alone. This work develops a CFD-informed machine-learning surrogate for spatial FD in a large-bore diesel engine, based on a Wärtsilä W20 injector and representative engine conditions. A fully coupled injector–spray–engine CFD framework under engine-like RCCI inert conditions determines the needle-lift profile and resolves the combined effects of injector geometry, needle dynamics, and operating conditions on in-cylinder flow, capturing physical phenomena not reproducible by isolated free-spray simulations. A high-fidelity database is generated using Latin Hypercube Sampling, from which FD is extracted at 15 CAD before top dead center within an annular multi-zone (MZ) representation consistent with reduced-order combustion models. A multi-output Random Forest (RF) surrogate, augmented with uncertainty-driven active learning, is trained to predict the complete spatial FD vector. Prediction errors are higher near the combustion chamber core than in liner-adjacent zones, reflecting stronger nonlinear coupling and localized data sparsity. To address this, four additional CFD samples are selected from regions of maximum predictive uncertainty and incorporated into the training dataset. This targeted enrichment markedly improves surrogate performance, reducing mean absolute error (MAE) under worst-case input conditions. Although localized error amplification persists in a few zones, these regions are systematically identified and can be mitigated through further adaptive sampling using candidates proposed by the updated surrogate. Convergence of the active-learning framework is assessed using mean MAE, worst-zone MAE, global L1 error, and ensemble-based predictive uncertainty, ensuring robust and consistent accuracy across the design space. The framework integrates CFD-resolved physics, machine-learning surrogates, uncertainty quantification, and adaptive sampling, providing a scalable and physically consistent approach for efficient FD prediction in advanced engines.