| description abstract | Abstract. This study developed a deep learning architecture to predict combustion characteristics and exhaust emissions using a prior soot–NOx coupled mechanism. Three-dimensional combustion simulations quantified the effects of operating conditions (ambient temperature, ambient pressure, exhaust gas recirculation (EGR) rate, fuel injection timing, and fuel injection quantity) on combustion and emissions. Increased ambient temperature, ambient pressure, fuel injection quantity, and advanced fuel injection timing advanced combustion phasing, raising in-cylinder pressure and rate of heat release while reducing ignition delay (most notably by 3.14 deg when increasing fuel injection quantity from 8 mg to 20 mg). Conversely, a higher exhaust gas recirculation rate (up to 20%) prolonged ignition delay by a maximum of 0.99 deg and weakened combustion intensity due to dilution and thermal effects. NOx emissions depended heavily on peak temperature, pressure, and residence time (e.g., increasing fuel injection quantity raised NOx by 99.3%), while soot was sensitive to local oxygen and mixture quality. The characteristic soot–NOx trade-off was confirmed, where applying EGR reduced NOx by 52.0% but increased soot by 99.0%. Based on 0D analysis, a long short-term memory model was selected for its superior temporal learning capabilities. Applied to the three-dimensional dataset (1200 time-series sequences), the long short-term memory model achieved an R2 > 0.96 for combustion characteristics and R2 > 0.98 for emission formation (NOx and soot) and oxidation processes. These results indicate that the model effectively captures the nonlinear relationships within the data. | |
