Condition Setup

Set Up Conditions for a PINA Problem

[!IMPORTANT] Read RULES.md before using this skill — it applies to all skills. This is a sub-skill of create-problem. Load the entry-point skill first.

Use this skill to define Condition objects that bind data, equations, or time-series windows to the problem.

Step 1 — Determine the condition type

Three kinds of conditions exist in PINA:

Step 2 — Data types (data-driven only)

If the problem is data-driven, ask:

What data type are you using?

Available data types for Condition(input=..., target=...):

• LabelTensor / torch.Tensor — standard tensor data (most common)
• Graph — PINA's built-in graph structure (from pina import Graph)
• Data — PyTorch Geometric Data object (from torch_geometric.data import Data)

All three types are accepted directly as input/target.

Step 3 — Build conditions

Physics-on-domain

Conditions map domain names (sampled later via discretise_domain) or explicit point tensors to equations:

from pina import Condition

# Option 1: reference a domain by name (sampled later)
conditions = {
    "boundary": Condition(domain="boundary", equation=FixedValue(0.0)),
    "interior": Condition(domain="D", equation=Equation(my_pde)),
}

# Option 2: provide explicit points
conditions = {
    "data_pde": Condition(input=points_tensor, equation=Equation(my_pde)),
}

Data-driven (supervised)

conditions = {
    "data": Condition(input=input_tensor, target=target_tensor),
}

Time series forecasting

If the user has time series data, ask whether they want standard supervised or time-series conditions:

from pina import Condition

# Standard supervised
Condition(input=ts_tensor, target=target_tensor)

# Time series (input is 3D: [batch, n_windows, features])
Condition(
    input=ts_tensor,
    n_windows=10,
    unroll_length=5,
    randomize=True,
)

# Graph time series
Condition(
    input=graph_ts_data,
    n_windows=10,
    unroll_length=5,
    key="some_key",
)

Parameters:

• n_windows — number of rolling windows
• unroll_length — prediction horizon per window
• randomize — shuffle window order
• key — key for graph time series data

Step 4 — Integrate with the problem class

Conditions become a class-level dict on the problem:

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

    domains = {
        "D": spatial_domain,
        "boundary": spatial_domain.partial(),
    }

    conditions = {
        "boundary": Condition(domain="boundary", equation=FixedValue(0.0)),
        "D": Condition(domain="D", equation=Equation(my_pde)),
    }

Checklist

• For data-driven: confirmed data type (LabelTensor, torch.Tensor, Graph, or PyG Data)
• For data-driven: input_variables is a list[str] naming the inputs
• For time series: n_windows, unroll_length, and optional key are set correctly
• Each Condition uses valid keyword arguments:
  • Condition(domain=..., equation=...) for physics-on-domain
  • Condition(input=..., equation=...) for physics-on-points
  • Condition(input=..., target=...) for data-driven
  • Condition(input=..., n_windows=..., unroll_length=...) for time series
• domains dict has an entry for every domain name used in conditions