xarray-simlab provides a framework for easily building custom computational models from a set of modular components (i.e., Python classes), called processes.
The framework handles issues that scientists who are developing models should not care too much about, like the model interface and the overall workflow management. Both are automatically determined from the succint, declarative-like interfaces of the model processes.
Notably via its xarray extension, xarray-simlab has already deep integration with the SciPy / PyData stack. Next versions will hopefully handle other technical issues like command line integration, interactive visualization and/or running many simulations in parallel, e.g., in the context of sensitivity analyses or inversion procedures.
xarray-simlab is being developped with the idea of reducing the gap between the environments used for building and running computational models and the ones used for processing and analyzing simulation results. If the latter environments become more powerful and interactive, progress has still to be done for the former ones.
xarray-simlab also encourages building new models from re-usable sets of components in order to avoid reinventing the wheel. In many cases we want to customize existing models (e.g., adding a new feature or slightly modifying the behavior) instead of building new models from scratch. This modular framework allows to do that with minimal effort. By implementing models using a large number of small components that can be easily plugged in/out, we eliminate the need of hard-coding changes that we want to apply to a model, which often leads to over-complicated code and interface.
The design of this tool is thus mainly focused on both fast model development and easy, interactive model exploration. Ultimately, this would optimize the iterative back-and-forth process between ideas that we have on how to model a particular phenomenon and insights that we get from the exploration of model behavior.
Sources of inspiration¶
xarray-simlab leverages the great number of packages that are part of the Python scientific ecosystem. More specifically, the packages below have been great sources of inspiration for this project.
- xarray: xarray-simlab actually provides an xarray extension for setting and running models.
- attrs: a package that allows writing Python classes without boilerplate. xarray-simlab uses and extends attrs for writing processes as succinct Python classes.
- luigi: the concept of Luigi is to use Python classes as re-usable units that help building complex workflows. xarray-simlab’s concept is similar, but here it is specific to computational (numerical) modeling.
- django (not really a scientific package): the way that model
processes are designed in xarray-simlab has been initially inspired
from Django’s ORM (i.e., the
- param: another source of inspiration for the interface of processes (more specifically the variables that it defines).
- climlab: another python package for process-oriented modeling, which uses the same approach although having a slightly different design/API, and which is applied to climate modeling.
- landlab: like climlab it provides a framework for building model components but it is here applied to landscape evolution modeling. It already has a great list of components ready to use.
- dask: represents fine-grained processing tasks as Directed Acyclic Graphs (DAGs). xarray-simlab models are DAGs too, where the nodes are interdepent processes. In this project we actually borrow some code from dask for resolving process dependencies and for model visualization.