Physics-Informed Neural Network PINN ★
Definition
A family of methods that encodes PDE residuals into the loss function and parameterizes the solution itself with a neural network: no mesh, no labeled solution, trained purely from the governing equations and initial/boundary conditions.
What it is
A Physics-Informed Neural Network represents the solution of a PDE directly with a network u(x, t; theta), then minimizes a weighted sum of losses for the equation residual, the initial condition and the boundary condition. Derivatives come from automatic differentiation, so no mesh is needed. A PINN solves one PDE instance: change the parameters or the geometry and you must retrain.
Key points
- Loss = data term + residual term + boundary term. Residuals are evaluated by AD at random collocation points.
- It can train without any labeled solution: governing equations plus conditions are enough, which matters when measurements are scarce.
- The solution is a closed-form, continuously differentiable function, so point evaluation and derivatives are cheap — ideal for inverse problems.
- Training is delicate: wildly different loss magnitudes need adaptive weights or curriculum learning.
Common pitfalls / boundaries
- A PINN solves one equation; to learn the solution operator for a family of equations, look at neural operators instead.
- Accuracy is usually below mature FEM/FVM solvers. The selling points are mesh-free geometry handling, data coupling and inverse problems — not raw accuracy.
- On high-frequency and multi-scale problems PINNs suffer from spectral bias and converge very slowly.