PEtab SciML - Scientific Machine Learning Format and Tooling ============================================================ PEtab SciML is a table-based data format for creating training (parameter estimation) problems for **scientific machine learning (SciML)** models that combine machine learning and mechanistic ordinary differential equation (ODE) models. .. warning:: **Beta Disclaimer**: this software is under active development and may contain bugs or instabilities. The PEtab SciML format is finalised and support for it has been implemented in PEtab importers, though not yet released. Documentation and utility functions are currently being added. Main features ------------- Extending the `PEtab format `_ for mechanistic ODE models, PEtab SciML provides a human readable, reproducible way to specify SciML training problems across diverse scenarios, in a format directly importable by downstream tools. The main features are: - **Flexible hybridization.** Machine learning (ML) and ODE models can be combined in three ways: (1) ML within the ODE dynamics (includes Neural ODEs), (2) ML in the observable/measurement model linking simulations to data, and (3) ML upstream of the ODE, mapping high-dimensional inputs (e.g., images) to ODE model parameters. - **Import across ecosystems.** PEtab SciML problems can be imported into state-of-the-art toolboxes for dynamic-model training in Julia (`PEtab.jl `_) and Python/JAX (`AMICI `_). - **Broad support for ML architectures.** A diverse set of ML architectures can be specified via an exchangeable PEtab SciML YAML format (supports export from PyTorch modules), or via importer-specific libraries (e.g., Lux.jl in PEtab.jl; Equinox in AMICI). - **Diverse model types.** All model features of the `PEtab format `_ are supported, like models with partial observability, multiple simulation conditions, diverse noise models, and/or events. - **Efficient training strategies.** With minimal user input, PEtab SciML problems can be rewritten at the PEtab abstraction level to be compatible with training strategies such as multiple shooting, curriculum learning, and regularization (e.g., of ML outputs). - **Thoroughly tested.** An extensive test suite ensures importers produce correct and consistent output. - **Linting and helpers.** The PEtab SciML Python library provides a linter and utility functions for creating common problem types (e.g., Neural ODEs) and transformations (e.g., rewriting a PEtab problem for multiple-shooting training). Installation ------------ The PEtab SciML Python3 helper library can be installed with: .. code-block:: bash pip install petab-sciml How to read the documentation ----------------------------- If you are new to PEtab SciML, start with the :doc:`Getting Started tutorial `. It is a prerequisite for the How-to guides, which cover different model scenarios (e.g., Neural ODEs, ML model upstream of the ODE). For a complete description of all options when defining a SciML problem, see the :doc:`Format specification `. Why a SciML data format? ------------------------ There are several technical challenges with training SciML models. Firstly, coding a correct and performant loss/objective function is non-trivial, especially for more complex scenarios like models with events/callbacks or partial observability. Secondly, efficient training requires gradients computed via automatic differentiation (AD). Although frameworks like JAX, PyTorch and the Julia SciML ecosystem support AD-compatible code, it can still be challenging to write performant code. Similarly, porting models between frameworks to leverage framework-specific benefits is time-consuming because each AD backend uses different syntax. As a result, training setups are often hard-coded to a single framework. This undermines reusability, both for extending prior work and for creating benchmark collections that can run across different toolboxes. PEtab SciML solves these problems by providing a human-writable, tabular, language-independent specification of the training objective. This further enables generic importers in different ecosystems (e.g., JAX, Julia SciML) to generate correct, performant training objectives without ad-hoc, per-project engineering. Together, this improves exchangeability and overall reproducibility. Getting Help with and Extending PEtab SciML ------------------------------------------- If you run into problems: - Check the :doc:`Troubleshooting ` section of the documentation. - If you encounter unexpected behavior or a bug, please open an `issue `_ on GitHub If PEtab SciML is missing a feature you need, please open an `issue `_ on GitHub