PINA Development Skill

Expert guidance for Physics-Informed Neural Networks (PINNs) and Scientific Machine Learning with PINA.

What is PINA?

PINA (Physics-Informed Neural networks for Advanced modeling) is a PyTorch-based library for solving partial differential equations (PDEs) using neural networks. It combines:

• Physics-Informed Neural Networks (PINNs): Solve forward and inverse PDE problems
• Neural Operators: FNO, DeepONet for operator learning
• Data-Driven Modeling: Supervised learning with physics constraints
• Reduced Order Modeling: POD-NN for efficient simulations

Built on: PyTorch, PyTorch Lightning, PyTorch Geometric

Core Workflow

Every PINA project follows these 4 steps:

from pina import Trainer
from pina.problem import SpatialProblem
from pina.solver import PINN
from pina.model import FeedForward

# Step 1: Define Problem
problem = MyProblem()
problem.discretise_domain(n=100, mode="grid")

# Step 2: Design Model
model = FeedForward(input_dimensions=1, output_dimensions=1, layers=[64, 64])

# Step 3: Define Solver
solver = PINN(problem, model)

# Step 4: Train
trainer = Trainer(solver, max_epochs=1000, accelerator='gpu')
trainer.train()

Simple ODE Example

from pina.problem import SpatialProblem
from pina.domain import CartesianDomain
from pina.condition import Condition
from pina.equation import Equation, FixedValue
from pina.operator import grad
import torch

def ode_equation(input_, output_):
    """PDE residual: du/dx - u = 0"""
    u_x = grad(output_, input_, components=["u"], d=["x"])
    u = output_.extract(["u"])
    return u_x - u

class SimpleODE(SpatialProblem):
    output_variables = ["u"]
    spatial_domain = CartesianDomain({"x": [0, 1]})

    domains = {
        "x0": CartesianDomain({"x": 0.0}),  # Boundary
        "D": CartesianDomain({"x": [0, 1]})  # Interior
    }

    conditions = {
        "bound_cond": Condition(domain="x0", equation=FixedValue(1.0)),
        "phys_cond": Condition(domain="D", equation=Equation(ode_equation))
    }

    def solution(self, pts):
        """Analytical solution for validation."""
        return torch.exp(pts.extract(["x"]))

problem = SimpleODE()

Models

FeedForward Networks

from pina.model import FeedForward

# Basic network
model = FeedForward(
    input_dimensions=2,
    output_dimensions=1,
    layers=[64, 64, 64],  # Hidden layers
    func=torch.nn.Tanh   # Activation function
)

# Alternative activations
model = FeedForward(
    input_dimensions=1,
    output_dimensions=1,
    layers=[100, 100, 100],
    func=torch.nn.Softplus  # or torch.nn.SiLU
)

See Custom Models Reference for advanced architectures including:

• Hard constraints
• Fourier feature embeddings
• Periodic boundary embeddings
• POD-NN
• Graph neural networks

See Neural Operators Reference for operator learning with FNO, DeepONet, and more.

PINN Solver

from pina.solver import PINN
from pina.optim import TorchOptimizer
import torch

pinn = PINN(
    problem=problem,
    model=model,
    optimizer=TorchOptimizer(torch.optim.Adam, lr=0.001)
)

See Advanced Solvers Reference for:

• Self-Adaptive PINN (SAPINN)
• Supervised Solver
• Custom solvers
• Training strategies

Training

Basic Training

from pina import Trainer
from pina.callbacks import MetricTracker

# Discretize domain
problem.discretise_domain(n=1000, mode="random", domains="all")

# Create trainer
trainer = Trainer(
    solver=pinn,
    max_epochs=1500,
    accelerator="cpu",  # or "gpu"
    enable_model_summary=False,
    callbacks=[MetricTracker()]
)

# Train
trainer.train()

Training Configuration

trainer = Trainer(
    solver=solver,
    max_epochs=1000,
    accelerator="gpu",
    devices=1,
    batch_size=32,
    gradient_clip_val=0.1,  # Gradient clipping
    callbacks=[MetricTracker()]
)
trainer.train()

Testing

# Test the model
test_results = trainer.test()

# Manual evaluation
with torch.no_grad():
    test_pts = problem.spatial_domain.sample(100, "grid")
    prediction = solver(test_pts)
    true_solution = problem.solution(test_pts)
    error = torch.abs(prediction - true_solution)

Domain Discretization

Sampling Modes

# Grid sampling (uniform points)
problem.discretise_domain(n=100, mode="grid", domains=["D", "x0"])

# Random sampling (Monte Carlo)
problem.discretise_domain(n=1000, mode="random", domains="all")

# Latin Hypercube Sampling
problem.discretise_domain(n=500, mode="lh", domains=["D"])

# Manual sampling
pts = problem.spatial_domain.sample(256, "grid", variables="x")

Best Practice: Start with grid for testing, use random/LH for training with more points.

Visualization

import matplotlib.pyplot as plt

@torch.no_grad()
def plot_solution(solver, n_points=256):
    # Sample points
    pts = solver.problem.spatial_domain.sample(n_points, "grid")

    # Get predictions
    predicted = solver(pts).extract("u").detach()
    true = solver.problem.solution(pts).detach()

    # Plot comparison
    fig, axes = plt.subplots(1, 3, figsize=(15, 5))

    axes[0].plot(pts.extract(["x"]), true, label="True", color="blue")
    axes[0].set_title("True Solution")
    axes[0].legend()

    axes[1].plot(pts.extract(["x"]), predicted, label="PINN", color="green")
    axes[1].set_title("PINN Solution")
    axes[1].legend()

    diff = torch.abs(true - predicted)
    axes[2].plot(pts.extract(["x"]), diff, label="Error", color="red")
    axes[2].set_title("Absolute Error")
    axes[2].legend()

    plt.tight_layout()
    plt.show()

See Visualization Reference for comprehensive plotting techniques.

Best Practices

1. Start Simple

# Begin with small network
model = FeedForward(input_dimensions=2, output_dimensions=1, layers=[20, 20])

# Gradually increase complexity
model = FeedForward(input_dimensions=2, output_dimensions=1, layers=[64, 64, 64])

2. Monitor Losses

from pina.callbacks import MetricTracker

trainer = Trainer(
    solver=pinn,
    max_epochs=1000,
    callbacks=[MetricTracker(["train_loss", "bound_cond_loss", "phys_cond_loss"])]
)

3. Two-Phase Training

# Phase 1: Rough solution (high LR)
pinn = PINN(problem, model, optimizer=TorchOptimizer(torch.optim.Adam, lr=0.01))
trainer = Trainer(pinn, max_epochs=500)
trainer.train()

# Phase 2: Refinement (low LR)
pinn.optimizer.param_groups[0]['lr'] = 0.001
trainer = Trainer(pinn, max_epochs=1500)
trainer.train()

MLflow Integration

Track PINA experiments with MLflow for reproducibility and comparison:

import mlflow
from pina import Trainer
from pina.solver import PINN

# Set experiment
mlflow.set_experiment("pina-poisson-solver")

with mlflow.start_run(run_name="baseline"):
    # Log hyperparameters
    mlflow.log_params({
        "layers": [64, 64, 64],
        "activation": "Tanh",
        "learning_rate": 0.001,
        "n_points": 1000,
        "epochs": 1500
    })

    # Setup and train
    problem.discretise_domain(n=1000, mode="random")
    trainer = Trainer(solver, max_epochs=1500)
    trainer.train()

    # Log final metrics
    mlflow.log_metric("final_loss", trainer.callback_metrics["train_loss"])

    # Log model
    mlflow.pytorch.log_model(solver.model, "pinn_model")

Marimo Dashboard Integration

Create interactive PINA dashboards with marimo:

import marimo as mo
from pina.solver import PINN

# UI controls for hyperparameters
layers = mo.ui.slider(1, 5, value=3, label="Hidden Layers")
neurons = mo.ui.slider(16, 128, value=64, step=16, label="Neurons/Layer")
lr = mo.ui.number(value=0.001, start=0.0001, stop=0.1, label="Learning Rate")

# Train button
train_btn = mo.ui.run_button(label="Train PINN")

# In another cell: run training when button clicked
if train_btn.value:
    model = FeedForward(
        input_dimensions=2,
        output_dimensions=1,
        layers=[neurons.value] * layers.value
    )
    # ... train and visualize

Using context7 for Documentation

Query up-to-date PINA documentation directly:

# context7 Library ID (no resolve needed):
# - /mathlab/pina (official docs, 2345 snippets)

# Example: query-docs("/mathlab/pina", "FeedForward model parameters")

When to Use This Skill

✅ Use PINA when:

• Solving PDEs with neural networks
• Need to incorporate physics constraints
• Working with inverse problems
• Building neural operators (FNO, DeepONet)
• Reduced order modeling
• Scientific ML research

❌ Don't use PINA when:

• Pure data-driven tasks (use standard PyTorch)
• Not dealing with differential equations
• Need classical numerical solvers (FEM, FVM)

Reference Documentation

Detailed documentation organized by topic:

• Problem Types: ODE, Poisson, Wave, Inverse problems, custom equations
• Neural Operators: FNO, DeepONet, Kernel Neural Operator
• Custom Models: Hard constraints, Fourier features, periodic embeddings, POD-NN, GNNs
• Advanced Solvers: SAPINN, supervised solver, custom solvers, training strategies
• Visualization: Plotting techniques, error analysis, animations

Complete Examples

Ready-to-run example scripts:

• Poisson 2D: Complete 2D Poisson equation solver with visualization
• Wave Equation: Time-dependent wave equation with animations
• FNO Example: Fourier Neural Operator for operator learning
• Inverse Problem: Learn unknown parameters from data

Resources

• Documentation: https://mathlab.github.io/PINA/
• GitHub: https://github.com/mathLab/PINA
• Paper: https://joss.theoj.org/papers/10.21105/joss.04813
• Tutorials: https://github.com/mathLab/PINA/tree/master/tutorials