Introduction

In anemoi-graphs, graphs are built using a recipe, which is a YAML file that specifies the nodes, edges, and attributes of an encoder-processor-decoder graph. The recipe file is used to create the graphs using the command-line tool.

The main components of the recipe file are the definition of the nodes, the edges, and each of these can optionally include attributes. Here we give an overview of these specifications; more details on the available options for each are given in the following pages.

Nodes

The nodes are defined in the nodes section of the recipe file. The keys are the names of the sets of nodes that will later be used to build the connections. Each nodes configuration must include a node_builder section describing how to define the nodes, and can optionally include node attributes.

A simple example of the nodes definition, for two sets of nodes, looks like this:

nodes:
  data:
    node_builder:
      _target_: anemoi.graphs.nodes.ZarrDatasetNodes
      dataset: /path/to/dataset.zarr
  hidden:
    node_builder:
      _target_: anemoi.graphs.nodes.TriNodes
      resolution: 5

In this example, two sets of nodes are defined, which have been named data and hidden (these names are later used to define edges). The data nodes have been defined based on coordinates specified in a zarr file and the hidden nodes are specified based on a triangular refined icosahedron algorithm.

Several methods are currently supported to create your nodes. You can use indistinctly any of these to create your data or hidden nodes:

• From anemoi dataset

• From NPZ file

• Reduced Gaussian grid

• Triangular Mesh with ICON Topology

• From text file

• From latitude & longitude coordinates

• Triangular refined Icosahedron

• Hexagonal refined Icosahedron

• HEALPix Nodes

In addition to the node_builder section, the nodes configuration can contain an optional section to define additional node attributes (weights, mask, …). For example, the weights can be used to define the importance of each node in the loss function, or the masks can be used to build connections only between subsets of nodes. See the attributes page for more details.

Edges

The edges section in the recipe file defines the edges between the nodes through which information will flow. These connections are defined between pairs of nodes sets (source and target, specified by source_name and target_name). There are several methods to build these edges, including cutoff (CutOffEdges) or nearest neighbours (KNNEdges).

For an encoder-processor-decoder graph you will need to build two sets of edges. The first set of edges will connect the data nodes with the hidden nodes to encode the input data into the latent space, normally referred to as the encoder edges and represented here by the first element of the edges section. The second set of edges will connect the hidden nodes with the data nodes to decode the latent space into the output data, normally referred to as decoder edges and represented here by the second element of the edges section.

Graphically, the encoder-processor-decoder setup looks like this:

The corresponding recipe file chunk is as follows:

edges:
  # A) Encoder connections
  - source_name: data
    target_name: hidden
    edge_builders:
    - _target_: anemoi.graphs.edges.CutOffEdges
      cutoff_factor: 0.7
  # B) Decoder connections
  - source_name: hidden
    target_name: data
    edge_builders:
    - _target_: anemoi.graphs.edges.KNNEdges
      num_nearest_neighbours: 3

In this example, the encoder edges are defined based on the cut-off radius, whereas the decoder nodes use a k-nearest neighbours algorithm. Available methods for defining edges are:

• Cut-off radius

• K-nearest neighbours

• Multi-scale connections

As with the nodes, the edges can have additional attributes - see the attributes page for more details.