PINNForge

Solve partial differential equations in 3 lines of code.

PINNForge is a lightweight wrapper around PyTorch that handles the boilerplate code of Physics-Informed Neural Networks so you can focus on the physics.

Install

pip install pinnforge

Quick Start

from pinnforge import solve_pde

result = solve_pde("heat", epochs=5000)
print(f"Relative L2 error: {result['metrics']['relative_l2_error']:.2e}")
# Relative L2 error: 3.42e-02

The library automatically builds the network, generates training points, trains, validates, and returns metrics.

Why PINNForge?

Setting up a PINN in the standard tooling requires 25+ lines of boilerplate before you see your first result. PINNForge does it in 3.

Side-by-side: solving the 1D Heat equation

from pinnforge import solve_pde

result = solve_pde("heat", epochs=5000)
print(result['metrics']['relative_l2_error'])

import deepxde as dde
import numpy as np

def pde(x, y):
    dy_t = dde.grad.jacobian(y, x, i=0, j=1)
    dy_xx = dde.grad.hessian(y, x, i=0, j=0)
    return dy_t - dy_xx

geom = dde.geometry.Interval(0, 1)
timedomain = dde.geometry.TimeDomain(0, 1)
geomtime = dde.geometry.GeometryXTime(geom, timedomain)

def boundary(x, on_boundary):
    return on_boundary

def initial(x):
    return np.sin(np.pi * x[:, 0:1])

bc = dde.icbc.DirichletBC(geomtime, lambda x: 0, boundary)
ic = dde.icbc.IC(geomtime, initial, lambda _, on_initial: on_initial)

data = dde.data.TimePDE(
    geomtime, pde, [bc, ic],
    num_domain=2000, num_boundary=100, num_initial=100,
)

net = dde.nn.FNN([2] + [64] * 3 + [1], "tanh", "Glorot normal")
model = dde.Model(data, net)
model.compile("adam", lr=1e-3)
model.train(epochs=15000)
model.compile("L-BFGS")
losshistory, train_state = model.train()

# Then write validation and plotting code yourself

Feature comparison

Features that PINNForge has that DeepXDE and PINNs-Torch do not:
    -One-line solve API
    -Symbolic PDE definition (SymPy)
    -Automatically generated validation metrics and plots
    -Built in numerical solvers

However, DeepXDE is multi-backend, is better for 3D+ problem, and tested since 2019 which are advantages over PINNForge. PINNForge trades depth for ease of use and time-to-result. If you need complex PDEs or 3D problems, use DeepXDE. If you want to try an idea in 5 minutes with minimal code, use PINNForge.

What's Included

• solve_pde(): one-line API for common PDEs (Heat, Burgers, Wave)
• AutoPINN: configurable solver with auto-selected hyperparameters
• PINN: the underlying neural network module
• SymbolicPDE: define custom PDEs using SymPy math notation
• NumericalSolver: finite-difference reference solutions
• ExperimentLogger: automatic training metrics and checkpoint logging

Custom PDEs with SymPy

For problems not in the built-in registry, define them symbolically:

import sympy as sp
from pinnforge import AutoPINN

x, t, u = sp.symbols('x t u')
equation = sp.diff(u, t) + u * sp.diff(u, x) - 0.01 * sp.diff(u, x, 2)

solver = AutoPINN.from_symbolic(equation, [x, t], u)
result = solver.solve()
print(result['metrics'])

Known Limitations (v0.1.0)

This is an early release. Some limitations are:

• Fourier feature embeddings are implemented but disabled by default. They currently cause training collapse on test problems (error jumps from 1e-2 to 1.2). Help wanted: see #1.
• Adaptive loss weighting is implemented but disabled by default. Same reason as Fourier.
• Adaptive activation (SA-PINN) is disabled by default. Interacts badly with Fourier features.
• Burgers equation validation returns {}. The analytical solution currently in the code is the Heat equation's solution, not Burgers'. Burgers has no simple solution for this initial condition. See #2.
• Only 1D problems are supported. Multi-dimensional PDEs are planned for v0.2.0.
• No CLI yet. Everything is Python-only.

Contributing

This is a new project and there are several tasks that would help a lot. If you're looking for a place to contribute to SciML tooling, help is greatly appreciated

Good First Issues

Larger Contributions Welcome

How to Contribute

1. Fork the repo
2. Create a feature branch (git checkout -b fix-fourier-features)
3. Make your changes
4. Run pytest tests/ -v and make sure everything passes
5. Open a Pull Request

See CONTRIBUTING.md for the full guide. First-time contributors are very welcome — if you're unsure where to start, open an issue and ask.

Tech Stack

• Python 3.8+
• PyTorch
• SymPy
• NumPy / SciPy

License

MIT — see LICENSE.

Acknowledgments

Built on the original Physics-Informed Neural Networks formulation by Raissi, Perikaris, and Karniadakis (2019). Inspired by the usability goals of DeepXDE but with a focus on time-to-first-result.