Design and optimization of cam profile for boxer engine via a novel deep learning framework

Improving the thermal efficiency of the boxer engine (BE) remains a persistent challenge in its development. The valve train system significantly constrains the achievable thermal efficiency, while conventional optimization approaches are computationally time-consuming and exhibit inherent limitations in addressing this issue. In this study, we propose attention-enhanced residual network (AER-Net) to predict the cam kinematics of valve train system across a wide range of cam profile design and BE parameters. The AER-Net architecture integrates multi-head self-attention with convolutional neural network to capture both local geometric features and global dependencies. The trained AER-Net model achieves an ARE below 5% and an R 2 greater than 97.71% respectively, indicating its effectiveness in efficiently identifying optimal cam profiles under strict kinematic constraints. The final optimal cam profile obtained from AER-Net increases lift by 0.26 mm and the fullness coefficient by 0.0021. Computationally, AER-Net can evaluate 10,000,000 profiles within 120 min, whereas traditional methods require about 15 days to evaluate only 500 profiles, demonstrating a substantial reduction in computational time.