DOI: 10.30939/ijastech..1935208 ISSN: 2587-0963

A Physics-Constrained Neural Network Framework for Surrogate Modeling and Multi-Objective Optimization of an RCCI Engine

Muhammet Emre Sanci
Show PDF Cite
Reactivity Controlled Compression Ignition (RCCI) engines offer a promising pathway toward simultaneous reduction of NOx, soot, and fuel consumption; however, their inherently nonlinear input-output behavior makes systematic calibration challenging. This study proposes a physics-constrained neural network (PCNN) surrogate model for a single-cylinder, naturally aspirated RCCI engine fueled with isopropanol and n-heptane blends (compression ratio 17.5, bore × stroke 85 × 90 mm, displacement 510 cm³). The model maps six operator-controllable inputs, namely premixed fuel ratio, excess air ratio, engine speed, intake air temperature, injection advance, and fuel mass flow rate, to ten simultaneous engine outputs covering torque, brake-specific fuel consumption, CO, HC, CO₂, NO, soot, IMEP, peak pressure rise rate, and indicated thermal efficiency. A thermodynamic consistency constraint derived from the BSFC identity is embedded directly into the composite training loss, coupling the predicted torque and fuel consumption outputs through known engine speed and fuel flow inputs. The PCNN is benchmarked against Random Forest, Support Vector Regression, and a classical feedforward ANN under an identical 80/20 train–test protocol. The proposed PCNN achieved the highest average test R² of 0.9584 across all ten outputs, outperforming Random Forest, SVR, and ANN baselines with corresponding average R² values of 0.8664, 0.9332, and 0.8657, respectively. The PCNN delivered the best individual R² in nine of the ten target variables, while SVR slightly outperformed PCNN for CO prediction. The trained surrogate is subsequently employed within a multi-start Sequential Quadratic Programming framework to identify optimal RCCI operating conditions under four engineering scenarios: maximum performance, minimum emissions, balanced multi-objective operation, and maximum fuel economy. Results confirm that no single operating point simultaneously satisfies all objectives, and the proposed framework provides an efficient, physics-informed tool for navigating these trade-offs.

More from our Archive