Performing physics simulations using deep learning (DL) is a glorious goal in the physicists’ community that recently adopted machine learning for various research tasks. However, accuracy of the DL prediction and the time required for the neural net training are the two key issues that must be overcome for the widespread use of DL in a physics simulation. Field theoretic simulation (FTS) is one promising branch of simulations for the study of order-to-disorder transition of polymers and their formation of nanostructures. However, it is a computationally expensive tool, and it may take up to one million time-steps of simulation for the calculation of ensemble averages of thermodynamic quantities. In this work, we propose a new Langevin FTS (L-FTS) method which utilizes a deep neural network (DNN) that can be successively applied to find the saddle point of pressure field accurately. Major deep learning (DL) models for the semantic segmentation in the computer vision are adopted to construct the optimal DNN architecture. Our model utilizing atrous convolutions in parallel is computationally efficient and it can be reused after single training for simulations at similar parameters. With our open-source L-FTS library, up to a sixfold speed increase is achievable without sacrifice of accuracy.