Meeting regulatory and customer demands requires detailed powertrain calibration which can be expensive and time-consuming. There is often a reliance on mathematical optimization tools to convert experimental learnings into a final calibration. This work focuses on developing multiple neural network machine learning (ML) models which were trained on different test-train data splits of test-cell recorded steady-state medium-duty (MD) diesel engine data. The output data was used to develop engine actuator maps by utilizing a genetic algorithm (GA). The genetic algorithm contains a fitness function which was varied to target different combinations of low NOx and CO2 emissions. The input variables used for the ML model were engine speed, engine torque, fuel rail pressure, exhaust gas recirculation (EGR) valve command, main injection timing, and wastegate valve command. The output variables predicted were NOx mass flow rate, exhaust temperature, fuel flow rate, and dry intake mass flow rate. The ML models were used to predict cycle-averaged engine-out emissions and time-series predictions of all output variables for different transient drive cycles. The drive cycles used for this case were the Heavy-Duty Federal Test Procedure (HDFTP) transient cycle, the Non-Road Transient Cycle (NRTC), the Ramped Mode Cycle (RMC) and the newly proposed on-road Low-Load Cycle (LLC).