During the last decade, haptic has been a new emerged and interesting subject for many researchers, which can be is the most important technology for computer haptics, which means the process of calculating the force or tactile feedback to give the user a sense of touch or interaction with the virtual object. Smooth haptic feedback is an important task for haptic rendering with complex virtual objects. However, commonly the update rate of the haptic rendering may drop down during contact in complex scenarios because the high computational cost is required for physics-based dynamic simulation. If the haptic rendering is done at a lower update rate, it may cause discontinuous or unstable force feedback. Therefore, to implement smooth and accurate haptic rendering, the update rate of force calculation should be kept in a high and constant frequency. In the current master thesis, we propose a novel real-time method based on machine learning to calculate smooth and accurate haptic feedback in complex scenarios. The method consists of two phases: data generator module and designing appropriate artificial neural network. In the first phase, we proposed an automated data generator module which provides data required for learning procedures. There are three sets of data: position of the haptic tool (input data), feedback force resulting from virtual object and the haptic tool interaction (output data 1) and deformed shape of virtual object after interaction (output data 2). Before going to phase two, preprocessing analysis is performed on the data. Three steps were done: 1) All data normalize using the Z-Score function, 2) The input data are transformed from Cartesian coordinate to cylindrical coordinate, 3) By using Zernike polynomials, the dimension of data acquired from deformed shape reduced as it can make the training calculation more efficient. These three steps make the learning procedure easier and faster. In the second phase, two artificial neural networks are designed in order to estimate the feedback force and deformed shape, respectively. The hyperparameters are tuned due to the systems complexity and dimension of inputs and outputs. The result shows that the proposed method can provide smooth and accurate haptic force feedback at a high update rate for complex scenarios. Moreover, it can predict the shape of the deformed object after interaction with reasonable error. Keywords Haptic Rendering, Machine Learning, Artificial Neural Network, Deformable Objects, Zernike Polynomials