multifunctional

Handling multifunctional activities in the Brightway LCA software framework.

Installation

You can install multifunctional via [pip] from [PyPI]:

$ pip install multifunctional

It is also available on anaconda using mamba or conda at the cmutel channel:

mamba install -c conda-forge -c cmutel multifunctional

Usage

Multifunctional activities can lead to linear algebra problems which don't have exactly one solution. Therefore, we commonly need to apply a handling function to either partition such activities, or otherwise manipulate their data such that they allow for the creation of a square and non-singular technosphere matrix.

This library is designed around the following workflow:

Users create and register a multifunctional.MultifunctionalDatabase. Registering this database must include the database metadata key default_allocation, which refers to an allocation strategy function present in multifunctional.allocation_strategies.

import multifunctional
mf_db = multifunctional.MultifunctionalDatabase("emojis FTW")
mf_db.register(default_allocation="price")

Multifunctional process(es) are created and written to the MultifunctionalDatabase. A multifunctional process is any process with multiple "functional" edges, either outputs (e.g. products) and/or input (e.g. wastes). Each functional edge must be labelled functional=True. This labeling can be done manually or via a classification function.

mf_data = {
    ("emojis FTW", "😼"): {
        "type": "product",
        "name": "meow",
        "unit": "kg",
    },
    ("emojis FTW", "🐶"): {
        "type": "product",
        "name": "woof",
        "unit": "kg",
    },
    ("emojis FTW", "1"): {
        "name": "process - 1",
        "location": "somewhere",
        "exchanges": [
            {
                "functional": True,
                "type": "production",
                "input": ("emojis FTW", "😼"),
                "amount": 4,
                "properties": {
                    "price": 7,
                    "mass": 6,
                },
            },
            {
                "functional": True,
                "type": "technosphere",
                "input": ("emojis FTW", "🐶"),
                "amount": 6,
                "properties": {
                    "price": 12,
                    "mass": 4,
                },
            },
        ],
    }
}

Allocation can be done manually before writing the data to the database; if not done manually, it will be done automatically upon calling .write()

LCA calculations can then be done as normal. See dev/basic_example.ipynb for a simple example.

Substitution

You don't need to use library for substitution, that already works natively in Brightway. Just produce a product which another process also produces (i.e. has the same database name and code), and the production amount of the other process will be reduced as needed to meet the functional unit demand.

We will eventually have a plan for including substitution in parallel with allocation.

Classifying functional edges

There is currently no built-in functionality to determine if an edge is functional based on its attributes - instead we rely on the label functional being manually specified. You can write a function to iterate over datasets and label the functional edges in whatever fashion you choose.

Built-in allocation functions

multifunctional includes the following built-in property-based allocation functions:

• price: Does economic allocation based on the property "price" in each functional edge.
• mass: Does economic allocation based on the property "mass" in each functional edge.
• manual_allocation: Does allocation based on the property "manual_allocation" in each functional edge. Doesn't normalize by amount of production exchange.
• equal: Splits burdens equally among all functional edges.

Property-based allocation assumes that each functional edge has a properties dictionary, and this dictionary has the relevant key with a corresponding numeric value. For example, for price allocation, each functional edge needs to have 'properties' = {'price': some_number}.

Custom property-based allocation functions

To create new property-based allocation functions, add an entry to allocation_strategies using the function property_allocation:

import multifunctional as mf
mf.allocation_strategies['<label in function dictionary>'] = property_allocation(property_label='<property string>')

Additions to allocation_strategies are not persisted, so they need to be added each time you start a new Python interpreter or Jupyter notebook.

Custom single-factor allocation functions

To create custom allocation functions which apply a single allocation factor to all nonfunctional inputs and outputs, pass a function to multifunctional.allocation.generic_allocation. This function needs to accept the following input arguments:

• edge_data (dict): Data on functional edge
• node: An instance of multifunctional.MaybeMultifunctionalProcess
• strategy_label: An optional string to label the allocation strategy used

The custom function should return a number.

The custom function needs to be curried and added to allocation_strategies. You can follow this example:

import multifunctional as mf
from functools import partial

def allocation_factor(edge_data: dict, node: mf.MaybeMultifunctionalProcess) -> float:
   """Nonsensical allocation factor generation"""
   if edge_data.get("unit") == "kg":
      return 1.2
   elif "silly" in node["name"]:
      return 4.2
   else:
      return 7

mf.allocation_strategies['silly'] = partial(
   mf.generic_allocation,
   func=allocation_factor,
   strategy_label="something silly"
)

Other custom allocation functions

To have complete control over allocation, add your own function to allocation_strategies. This function should take an input of either multifunctional.MaybeMultifunctionalProcess or a plain data dictionary, and return a list of data dictionaries including the original input process. These dictionaries can follow the normal ProcessWithReferenceProduct data schema, but the result datasets need to also include the following:

• mf_parent_key: Integer database id of the source multifunctional process
• type: One of "readonly_process", "process", or "multifunctional"

Furthermore, the code of the allocated processes (mf_allocated_process_code) must be written to each functional edge (and that edge saved so this data is persisted). See the code in multifunctional.allocation.generic_allocation for an example.

Technical notes

Process-specific allocation strategies

Individual processes can override the default database allocation by specifying their own default_allocation:

import bw2data
node = bw2data.get(database="emojis FTW", code="1")
node["default_allocation"] = "mass"
node.save()

Specifying code values for allocated processes

When allocating a multifunctional process to separate monofunctional processes, we need to generate code values for each monofunctional process. This can be done by specifying desired_code for the functional exchange. See dev/basic_example.ipynb for a simple example.

When writing a multifunctional process, we need to create artificial edges to allocated processes which don't exist yet. You can't therefore directly specify the code of such an edge.

Separate product nodes

By default, the allocation function creates chimaera process+product nodes. However, we recommend distinguishing products and processes as best practice, and this is supported by multifunctional. You will need to create the product nodes yourself; they can be in the same multifunctional database, or in another database.

To create a functional link to a product node in the same database, you should specify an exchange input to the desired product. See dev/split_products.ipynb for a simple example. The product can be in the mutifunctional database, but doesn't have to be.

How does it work?

Recent Brightway versions allow users to specify which graph nodes types should be used when building matrices, and which types can be ignored. We create a multifunctional process node with the type multifunctional, which will be ignored when creating processed datapackages. However, in our database class MultifunctionalDatabase we change the function which creates these processed datapackages to load the multifunctional processes, perform whatever strategy is needed to handle multifunctionality, and then use the results of those handling strategies (e.g. monofunctional processes) in the processed datapackage.

We also tell MultifunctionalDatabase to load a new ReadOnlyProcessWithReferenceProduct process class instead of the standard Activity class when interacting with the database. This new class is read only because the data is generated from the multifunctional process itself - if updates are needed, either that input process or the allocation function should be modified.

Contributing

Contributions are very welcome. To learn more, see the Contributor Guide.

License

Distributed under the terms of the BSD 3 Clause license, multifunctional is free and open source software.

Issues

If you encounter any problems, please file an issue along with a detailed description.

Building the Documentation

You can build the documentation locally by installing the documentation Conda environment:

conda env create -f docs/environment.yml

activating the environment

conda activate sphinx_multifunctional

and running the build command:

sphinx-build docs _build/html --builder=html --jobs=auto --write-all; open _build/html/index.html