neurodiffeq

Why we included this project

Solving a differential equation numerically usually means laying down a grid and stepping through it, which gets awkward when you want the solution to plug straight into a PyTorch training loop. NeuroDiffEq takes a different route: you give it the equation plus its initial or boundary conditions, and it trains a neural network to approximate the solution directly. The API stays small, you define the system as a Python function, choose a network and condition type, and a solver hands back a callable solution object you can evaluate at any point. It covers ODE systems like Lotka-Volterra and 2D PDEs such as Laplace's equation, and it can also handle solution bundles and reverse problems, so it is not limited to textbook cases. The project is already used by research groups including Harvard's IACS, and teams working in PyTorch will find it a natural fit when they want a smooth, continuous approximation rather than a gridded numerical result.