Fuel economy improvement and stringent emission regulations worldwide require advanced air charging and combustion technologies, such as low temperature combustion, PCCI or HCCI combustion. Furthermore, NOx aftertreatment systems, like Selective Catalyst Reduction (SCR) or lean NOx trap (LNT), are needed to reduce vehicle tailpipe emissions. The information on engine-out NOx emissions is essential for engine combustion optimization, for engine and aftertreatment system development, especially for those involving combustion optimization, system integration, control strategies, and for on-board diagnosis (OBD).A physical NOx sensor involves additional cost and requires on-board diagnostic algorithms to monitor the performance of the NOx sensor. A control-oriented engine-out NOx model that meets throughput constraints is desirable for on-board applications as a virtual NOx sensor, which could be used in replacement of the physical sensor or run parallel as an analytical redundancy to detect potential in-range failure for the physical sensor.This paper presents the design of a virtual NOx sensor using neural networks combined with nonlinear dynamic system. NOx emission estimation with or without cylinder pressure information is investigated, and a correction to a baseline NOx model is proposed in order to compensate transient EGR responses. The developed virtual NOx sensor demonstrates a higher level of fidelity estimation when compared with actual NOx measurements.