Predicting end-of-process variables in internal combustion engines, such as brake-specific fuel consumption or pollutant emissions, is crucial for engine design decisions. However, errors in common multi-layer-perceptron-based artificial neural network models often match the magnitude of the expected fuel consumption improvements, potentially leading to incorrect decisions. This study introduces a hybrid model where artificial neural networks replace engine block elements, while the 1D gas circuit and turbocharger models are retained. To enhance metamodel accuracy, two modifications are proposed: incorporating a pulsating mass flow rate in the exhaust line to capture pulsating effects missing in mean-value engine models and using a hierarchical arrangement of several multi-layer perceptrons instead of a parallel arrangement. The pulsating mass flow rate approach improves the accuracy of all tested configurations by replicating pulsating effects from a detailed 1D engine model. Meanwhile, the hierarchical arrangement refines predictions of end-of-process variables, such as fuel consumption, by increasing the total layers, with a minimal trade-off in the accuracy of other variables. These findings are validated using a metamodel derived from a calibrated 1D engine model in GT-Suite. The proposed methods are expected to enhance the accuracy of data-driven artificial neural network approaches, contributing to more reliable engine design optimization.