In the past decades, the use of machine learning has significantly increased, particularly in the field of neural networks. In this article, a method to reconstruct the temperature field is proposed by taking the data at the boundary of a square slab. The number of virtual temperature sensors (VTS) has been selected for the training of the model, and the temperature field is reconstructed. In the first instance of this computational experimentation, the learning rate of the network was optimized with the maximum number of VTS. The learning rate 1e⁻⁵ was found to be suitable for neural network training. Further, this learning rate has been used to train the model with 24 and 16 VTS at the boundary. The number of VTS has been reduced to 16 to reconstruct the temperature field with acceptable accuracy. To check the robustness of the model, the temperature field prediction was performed with different boundary conditions by an already trained model.