The combustion process in Homogeneous Charge Compression Ignition Engines (HCCI) is one of the new methods of futuristic combustion technologies. Since there is no direct operator for the start of the combustion (SOC) of these engines, air-fuel mixture properties at the moment of entering the combustion chamber, specifies the ignition timing. In HCCI engines, the ignition timing is the most crucial factor in determining other engine operating characteristics such as power output, pollution, and fuel consumption. To control SOC, there should be an accurate predictive model based on the entering air-fuel mixture properties. The Artificial Neural Networks (ANN) approach can be considered as a solution with less computational costs than traditional physics-based modeling. In this investigation, a multi-input single-output model was developed for predicting the SOC of the HCCI engine for a wide range of engine operation. Three popular architectures namely the Nonlinear Autoregressive Network with Exogenous Inputs (NARXNET), Multi-Layer Perceptron (MLP) and Radial Basis Function (RBF) were used, for this purpose. The networks were trained using experimental data taken from a one-cylinder Ricardo engine. The network architecture was optimized using a Genetic Algorithm (GA) method. By using GA, the proposed networks also have the optimum network structures, improved model predictive behaviors, and simulation costs of the learning process. After optimization, the regression ratio between the outputs of MLP and the corresponding experimental data was increased from 0.8965 to 0.96166. This value was improved from 0.7623 to 0.83991 for RBF. By using GA, the time needed to train the NAM( was reduced from 3.12 s to 0.46 s. By comparing the model predictions with the experimental data, it was shown that the selected neural network architectures are powerful approaches for non-linear modeling the SOC of the HCCI engine.