anypinn 0.25.4

...

AnyPINN

Solve differential equations with Physics-Informed Neural Networks.
_{Modular. Training-agnostic. Inverse-problem-first.}

Most PINN libraries make you wire up every loss term, collocation grid, and training loop by hand before you see a single result. AnyPINN gives you a working experiment in one command and then lets you peel back every layer when you're ready.

🚀 Quick Start

The fastest way to start is the bootstrap CLI. It scaffolds a complete, runnable project interactively. Run it with uvx (ships with uv):

uvx anypinn create my-project

or with pipx:

pipx run anypinn create my-project

Run anypinn create --help to see all available flags and templates. For a full walkthrough covering project structure, configuration, training, and next steps, see the Getting Started guide.

👥 Who Is This For?

AnyPINN is built around progressive complexity. Start simple, go deeper only when you need to.

User	Goal	How
Experimenter	Run a known problem, tweak parameters, see results	Pick a built-in template, change config, press start
Researcher	Define new physics or custom constraints	Subclass Constraint and Problem, use the provided training engine
Framework builder	Custom training loops, novel architectures	Use anypinn.core directly, no Lightning required

💡 Examples

The examples/ directory has ready-made, self-contained scripts covering epidemic models, oscillators, predator-prey dynamics, and more, from a minimal ~80-line core-only script to full Lightning stacks. They're a great source of inspiration when defining your own problem.

🔬 Defining Your Own Problem

If you want to go beyond the built-in templates, here is the full workflow for defining a custom inverse problem.

1: Define the ODE

Implement a function matching the ODECallable protocol:

from torch import Tensor
from anypinn.core import ArgsRegistry

def my_ode(x: Tensor, y: Tensor, args: ArgsRegistry) -> Tensor:
    """Return dy/dx given current state y and position x."""
    k = args["k"](x)        # learnable or fixed parameter
    return -k * y           # simple exponential decay

2: Configure hyperparameters

from dataclasses import dataclass
from anypinn.problems import ODEHyperparameters

@dataclass(frozen=True, kw_only=True)
class MyHyperparameters(ODEHyperparameters):
    pde_weight: float = 1.0
    ic_weight: float = 10.0
    data_weight: float = 5.0

3: Build the problem

from anypinn.problems import ODEInverseProblem, ODEProperties

props = ODEProperties(ode=my_ode, args={"k": param}, y0=y0)
problem = ODEInverseProblem(
    ode_props=props,
    fields={"u": field},
    params={"k": param},
    hp=hp,
)

4: Train

import pytorch_lightning as pl
from anypinn.lightning import PINNModule

# With Lightning (batteries included)
module = PINNModule(problem, hp)
trainer = pl.Trainer(max_epochs=50_000)
trainer.fit(module, datamodule=dm)

# Or with your own training loop (core only, no Lightning)
optimizer = torch.optim.Adam(problem.parameters(), lr=1e-3)
for batch in dataloader:
    optimizer.zero_grad()
    loss = problem.training_loss(batch, log=my_log_fn)
    loss.backward()
    optimizer.step()

🏗️ Architecture

AnyPINN is split into four layers with a strict dependency direction: outer layers depend on inner ones, never the reverse.

graph TD
    EXP["Your Experiment / Generated Project"]

    EXP --> CAT
    EXP --> LIT

    subgraph CAT["anypinn.catalog"]
        direction LR
        CA1[SIR / SEIR]
        CA2[DampedOscillator]
        CA3[LotkaVolterra]
    end

    subgraph LIT["anypinn.lightning (optional)"]
        direction LR
        L1[PINNModule]
        L2[Callbacks]
        L3[PINNDataModule]
    end

    subgraph PROB["anypinn.problems"]
        direction LR
        P1[ResidualsConstraint]
        P2[ICConstraint]
        P3[DataConstraint]
        P4[ODEInverseProblem]
    end

    subgraph CORE["anypinn.core (standalone · pure PyTorch)"]
        direction LR
        C1[Problem · Constraint]
        C2[Field · Parameter]
        C3[Config · Context]
    end

    CAT -->|depends on| PROB
    CAT -->|depends on| CORE
    LIT -->|depends on| CORE
    PROB -->|depends on| CORE

anypinn.core: The Math Layer

Pure PyTorch. Defines what a PINN problem is, with no opinions about training.

• Problem: aggregates constraints, fields, and parameters. Provides training_loss() and predict().
• Constraint (ABC): a single loss term. Subclass it to express any physics equation, boundary condition, or data-matching objective.
• Field: MLP mapping input coordinates to state variables (e.g., t → [S, I, R]).
• Parameter: learnable scalar or function-valued parameter (e.g., β in SIR).
• InferredContext: runtime domain bounds and validation references, extracted from data and injected into constraints automatically.

anypinn.lightning: The Training Engine (optional)

A thin wrapper plugging a Problem into PyTorch Lightning:

• PINNModule: LightningModule wrapping any Problem. Handles optimizer setup, context injection, and prediction.
• PINNDataModule: abstract data module managing loading, config-driven collocation sampling, and context creation. Collocation strategy is selected via TrainingDataConfig.collocation_sampler ("random", "uniform", "latin_hypercube", "log_uniform_1d", or "adaptive").
• Callbacks: SMMA-based early stopping, formatted progress bars, data scaling, prediction writers.

anypinn.problems: ODE Building Blocks

Ready-made constraints for ODE inverse problems:

• ResidualsConstraint: ‖dy/dt − f(t, y)‖² via autograd
• ICConstraint: ‖y(t₀) − y₀‖²
• DataConstraint: ‖prediction − observed data‖²
• ODEInverseProblem: composes all three with configurable weights

anypinn.catalog: Problem-Specific Building Blocks

Drop-in ODE functions and DataModules for specific systems. See anypinn/catalog/ for the full list.

🤝 Contributing

See CONTRIBUTING.md for setup instructions, code style guidelines, and the pull request workflow.