This article presents a new methodology called deep Theory of FunctionalConnections (TFC) that estimates the solutions of partial differentialequations (PDEs) by combining neural networks with TFC. TFC is used totransform PDEs with boundary conditions into unconstrained optimizationproblems by embedding the boundary conditions into a "constrained expression."In this work, a neural network is chosen as the free function, and used tosolve the now unconstrained optimization problem. The loss function is taken asthe square of the residual of the PDE. Then, the neural network is trained inan unsupervised manner to solve the unconstrained optimization problem. Thismethodology has two major differences when compared with popular methods usedto estimate the solutions of PDEs. First, this methodology does not need todiscretize the domain into a grid, rather, this methodology randomly samplespoints from the domain during the training phase. Second, after training, thismethodology represents a closed form, analytical, differentiable approximationof the solution throughout the entire training domain. In contrast, otherpopular methods require interpolation if the estimated solution is desired atpoints that do not lie on the discretized grid. The deep TFC method forestimating the solution of PDEs is demonstrated on four problems with a varietyof boundary conditions.