The paper proposes deep neural networks architectures that are designed to allow for an efficient evaluation of differential operators. The approach is simple, original and likely to trigger further innovations in the coming years. It will also provide a baseline for future methods based on other approaches to solving the same type of problems.