A real-time approach has been developed and assessed to control BMEP (brake mean effective pressure) and MFB50 (crank angle at which 50% of fuel mass has burnt) in a Euro 6 1.6L GM diesel engine. The approach is based on the use of feed-forward ANNs (artificial neural networks), which have been trained using virtual tests simulated by a previously developed low-throughput physical engine model. The latter is capable of predicting the heat release and the in-cylinder pressure, as well as the related metrics (MFB50, IMEP - indicated mean effective pressure) on the basis of an improved version of the accumulated fuel mass approach. BMEP is obtained from IMEP taking into account friction losses. The low-throughput physical model does not require high calibration effort and is also suitable for control-oriented applications. However, control tasks characterized by stricter demands in terms of computational time may require a modeling approach characterized by a further lower throughput. To this aim, feed-forward NNs have been trained to predict MFB50 and BMEP using a large dataset of virtual tests generated by the well-calibrated low-throughput physical engine model. The real-time approach has also been applied to derive the start of injection of the main pulse and the injected fuel quantity to achieve specific targets of MFB50 and BMEP. The accuracy of the real-time approach has been assessed based on experimental data taken at GM-GPS (General Motors - Global Propulsion Systems) facilities and its computational time has been compared to that of the low-throughput physical engine model, at steady-state and transient conditions over the WLTP cycle.