Contributing emulators

This guide explains how to contribute new emulator models to AutoEmulate.

Emulator structure

All emulators in AutoEmulate are implemented as scikit-learn estimators, making them compatible with scikit-learn’s cross-validation, grid-search, and pipeline functionality. Have a look at the scikit-learn estimator developer guide for more details on how to implement a new emulator.

Note: Keep in mind when contributing emuulators that AutoEmulate doesn’t currently support time-series or spatial data.

Core Requirements

Each emulator class must:

1. Live in autoemulate/emulators/

2. Inherit from sklearn.base’s BaseEstimator and RegressorMixin

3. Implement the fit and predict methods

4. Include these additional methods/properties:

   • get_grid_params(): Returns a dictionary of parameter values for grid search over hyperparameters

   • model_name: Property that returns the emulator name (usually self.__class__.__name__)

   • _more_tags(): Defines emulator properties like multioutput support

Getting Started

The easiest way to create a new emulator is to:

1. Look at existing emulators in autoemulate/emulators/ as templates

2. Run the scikit-learn estimator tests tests/test_estimators.py early to catch any implementation issues

3. Add your own tests in tests/models/

Naming Conventions

The model_name property allows the emulator to be accessed with both long and short names:

• Long name: The class name (e.g., “RadialBasisFunctions”)

• Short name: Uppercase letters from long name (e.g., “rbf”)

Make sure your chosen class name:

• Doesn’t conflict with existing emulators

• Contains some uppercase letters for the short name

• Is descriptive of the emulation technique

Testing emulators

We use two types of tests:

1. Scikit-learn Test Suite: Add your emulator to tests/test_estimators.py to verify scikit-learn compatibility. Not all tests need to pass - use _more_tags() to skip incompatible tests. See the estimator tags overview for details.

2. Custom Tests: Add specific tests for your emulator in tests/models/ to verify its core functionality (e.g., validating end-to-end functionality of components such as parameter search etc).

Registering an emulator

After your emulator passes tests:

1. Add it to model_registry in autoemulate/emulators/__init__.py

2. Set is_core=False to make it available but not a default model

PyTorch emulators

PyTorch emulators require special handling:

1. Put the model architecture in autoemulate/emulators/neural_networks/

2. Put the main emulator class in autoemulate/emulators/

3. Use skorch for scikit-learn compatibility:

   • Create self.model_ as NeuralNetRegressor instance

   • Pass model architecture as first argument

   • Use self.model_ in fit and predict methods

See existing PyTorch emulators like conditional_neural_process.py for examples.