Reynolds-Averaged Navier Stokes (RANS) method is highly favoured for turbulent flow simulation due to its efficiency and relatively lower computational cost than other methods. This method requires an appropriate turbulence model to obtain an accurate turbulent flow solution. In this work, we develop three machine learning-based turbulence models and one of them is our proposed Feedforward Neural Network (FFNN) architecture. All of these models are used to calculate the Reynolds stress based on the flow properties from RANS data. These models are tested in several flow configurations where the RANS model as the conventional method struggles. This study demonstrates that our proposed neural network model surpasses the benchmark performance of the conventional k ? ? turbulence model and outperforms other machine learning-based turbulence models. The results show that the Reynolds stress field predicted by the proposed neural network closely resembles the Direct Numerical Simulation and Large Eddy Simulation. In this work, we also studied the effect of input feature selection on the model’s accuracy and performed feature importance analysis using SHAPley Additive Explanation (SHAP) that supports the result of our study of input feature effect.