select-solver

Select a Solver for a PINA Problem

[!IMPORTANT] Read RULES.md before using this skill — it applies to all skills.

Use this skill to pick the correct PINA solver — or design a custom one — based on your problem type, conditions, and training requirements.

PINA solvers are driven by conditions (data format) and model count (single vs ensemble). The condition types in your problem.conditions dictate which solver families are compatible; additional requirements (causality, gradient enhancement, adaptive weighting) narrow the choice further.

Step 1 — Understand the conditions

Start by identifying what kind of data your problem works with. Ask conversationally if the user hasn't already built their problem object:

• Time series? — Your conditions use n_windows and unroll_length. The data has a sequential structure and the model rolls out predictions step by step.
• Input → Target? — You have input points and known target values. This is standard supervised / regression data: Condition(input=..., target=...).
• PDE equations + domains? — Your conditions reference equations and domains/spatial points: Condition(domain=..., equation=...) or Condition(input=..., equation=...).
• Mixed? — Many physics-informed problems mix equation conditions (for the PDE interior) with target conditions (for boundary data or observations).

You can inspect an existing problem object directly:

for name, cond in problem.conditions.items():
    print(name, type(cond).__name__)

If the user hasn't defined conditions yet, ask if you can help setting up the problem.

Step 2 — Let the conditions drive the choice

Once the condition types are clear, map them to solver families. This is a decision tree, not a menu — work through it with the user.

Time series conditions

If every condition is (or includes) a TimeSeriesCondition:

→ AutoregressiveSingleModelSolver  (one model)
→ AutoregressiveEnsembleSolver     (multiple models, averaged)

No other solver family handles TimeSeriesCondition.

Key parameters:

• eps — noise injected during unrolling (regularisation)
• unroll_length — number of steps per forward pass

Input → Target conditions only

If all conditions are InputTargetCondition (no equations, no domains):

→ SupervisedSingleModelSolver  (one model)
→ SupervisedEnsembleSolver     (N models with deep ensemble)

The Ensemble variant trains N independent copies and averages their predictions at inference.

Equation conditions (physics-informed)

If conditions include InputEquationCondition or DomainEquationCondition, you're in the physics-informed family. Now ask two follow-up questions:

1. How many models: one or multiple?

Multiple models trains N independent copies and averages predictions.

2. Any special training requirement?

For the single-model physics-informed case, ask the user about these. Each maps to a distinct solver:

Self-adaptive and competitive are not single-model solvers — they manage multiple optimisers internally. SelfAdaptive trains a weight network alongside the model; Competitive trains a discriminator.

Ensemble + special requirements? The PhysicsInformedEnsembleSolver does not layer gradient/causal/self-adaptive on top. For those combinations, create a custom solver (Step 4).

Step 3 — Assemble the solver

Once the solver is chosen, the construction pattern is consistent:

from pina.solver import PhysicsInformedSingleModelSolver
# or any other solver

solver = PhysicsInformedSingleModelSolver(
    problem=problem,
    model=model,
    learning_rate=0.001,
    # optional:
    # loss=torch.nn.MSELoss(),
    # weighting=my_weighting,
    # scheduler=torch.optim.lr_scheduler.CosineAnnealingLR(...),
    # batch_size=32,
)

Step 4 — Creating a custom solver

When no built-in solver matches your requirements, PINA's architecture makes it straightforward to compose one. The key building blocks are mixins — reusable training_step and _compute_condition_loss overrides.

When to create a custom solver

• You need a combination PINA doesn't provide (e.g., causal weighting + ensemble, or gradient-enhanced + competitive)
• Your loss computation has custom per-condition logic not covered by the existing _compute_condition_loss overrides
• You need a multi-model setup that isn't just an ensemble (e.g., two models exchanging information in a non-adversarial way)
• You need to override training_step for a non-standard optimiser loop

How PINA solvers are composed

Every solver is a class hierarchy:

BaseSolver
 ├── training/val/test_step
 ├── _compute_condition_loss
 ├── _prepare_condition_data
 ├── _regularize_condition_loss
 └── _loss_from_residual

SingleModelSolver(BaseSolver)
 └── automatic_optimization=True, single forward

MultiModelSolver(BaseSolver)
 └── automatic_optimization=False, multiple optimisers

EnsembleSolver(BaseSolver)          — uses MultiModelSolver internally
 └── automatic_optimization=False, averages N models

Mixins override specific methods:

Custom solver recipe

1. Pick a base: SingleModelSolver (auto-optim) or MultiModelSolver (manual optim for multiple optimisers) or EnsembleSolver (N-model avg).

2. Mix in behaviours by inheriting the mixins in the right order (mixins first so their method overrides take priority):

   from pina.solver import SingleModelSolver
   from pina.solver.mixin import GradientEnhancedMixin, PhysicsInformedMixin
   
   class MyCustomSolver(
       PhysicsInformedMixin,      # 1st — ensures autograd in val/test
       GradientEnhancedMixin,     # 2nd — adds gradient penalty
       SingleModelSolver,         # 3rd — base training loop
   ):
       def __init__(self, problem, model, **kwargs):
           super().__init__(problem=problem, model=model, **kwargs)
   

3. Override _compute_condition_loss if you need custom per-condition logic (e.g., different loss functions per condition, special weighting). The signature is:

   def _compute_condition_loss(self, condition, data, batch_idx):
       # data is a dict like {"input": tensor, "target": tensor}
       # condition is the Condition object
       ...
       return scalar_tensor
   

4. Override training_step only for fundamental changes to the optimiser loop (e.g., alternating min-max, multi-stage schedules):

   def training_step(self, batch, batch_idx):
       # Custom loop: zero_grad → compute loss → manual_backward → step
       ...
   

Example: Causal ensemble

from pina.solver import EnsembleSolver

class CausalEnsemblePINN(EnsembleSolver):
    """Ensemble of models with causal temporal weighting."""
    def _compute_condition_loss(self, condition, data, batch_idx):
        # Inject causal weighting logic here
        # (iterate over time segments, apply causal mask)
        ...

When guiding a user through creating a custom solver, discuss their specific need, identify which existing mixin or base class covers most of it, and then describe only the method they need to override. Avoid generating the full solver class unless the user explicitly asks for it.

Templates

Choosing from conditions

Use this pattern when the problem object already exists:

from pina.solver import (
    SupervisedSingleModelSolver,
    PhysicsInformedSingleModelSolver,
    AutoregressiveSingleModelSolver,
    PhysicsInformedEnsembleSolver,
)

condition_types = {type(c).__name__ for _, c in problem.conditions.items()}

if "TimeSeriesCondition" in condition_types:
    solver = AutoregressiveSingleModelSolver(problem=problem, model=model)
elif "InputTargetCondition" in condition_types and not \
     ("InputEquationCondition" in condition_types or
      "DomainEquationCondition" in condition_types):
    solver = SupervisedSingleModelSolver(problem=problem, model=model)
else:
    solver = PhysicsInformedSingleModelSolver(problem=problem, model=model)

Custom solver stub

from pina.solver import SingleModelSolver

class MySolver(SingleModelSolver):
    """Custom solver for specialised loss logic."""

    def _compute_condition_loss(self, condition, data, batch_idx):
        # Your custom loss computation
        # data keys: "input", "target" (if InputTargetCondition)
        #            "input", "equation" (if InputEquationCondition)
        ...
        return loss

    def training_step(self, batch, batch_idx):
        # Only if you need a custom optimiser loop
        ...

Checklist

• Condition types identified (time series / supervised / physics-informed / mixed)
• Condition-to-solver-family mapping discussed with the user
• Single-model vs ensemble decision made
• Special training requirements assessed (causality, gradient, adaptive, competitive)
• If custom solver: mixin composition pattern understood, only necessary overrides discussed
• Solver constructed with correct signature (problem, model/models, extra params)