The traditional Simulink model is computationally inefficient due to solving the system of coupled differential equations, while the existing data-driven approach suffers from dynamic response hysteresis and lack of real-time performance in diesel engine transient condition prediction. In this study, aiming at the need for High-Fidelity and Real-Time Interaction in Digital Twin modeling of diesel intake and exhaust systems, this paper proposes a lightweight surrogate modeling method based on Neural Ordinary Differential Equations. we innovatively integrate Simulink simulation data and continuous-time dynamics modeling theory to construct a multilayer perceptron parameterized differential equation model with system state variables (intake flow, exhaust flow, and exhaust enthalpy) and control variables (rotational speed) as inputs, and use an adaptive Runge-Kutta solver to realize continuous state prediction. The experimental results show that the proposed surrogate model has reduced the root mean square error and maximum absolute error on the test set and further improved the fitting accuracy compared with the Long Short-Term Memory and Residual Neural Network. The coefficient of determination under unseen conditions exceeds 0.98, which verifies its excellent generalization performance, and the simulation speed is 5 times faster than the Simulink model.