Neurals

ADELM assembles each parameter from one of three sources declared in the configuration. The source can be changed between runs without modifying the process code — the forward equations are the same regardless of where each parameter value comes from.

  • Prescribed — a fixed value, a PFT-weighted lookup, or a pedotransfer-derived quantity. The conventional approach used in most land surface models.

  • nn_global — a globally shared learned value: one scalar per parameter, optimized from observations and applied uniformly to all locations.

  • nn_feature_based — a spatially varying learned parameter: an attribute-to-parameter mapping from static environmental attributes to parameter values. Trained at observation sites, it is evaluated across the full spatial domain to produce a continuous parameter field.

Prescribed and learned sources can be freely combined within the same run.

See also

Neural Parameterization — technical details of nn_global and nn_feature_based.

Configuration — how to declare parameter sources in config.yaml.

Why observation-constrained parameterization?

Land surface parameters such as stomatal conductance or soil hydraulic properties cannot be measured directly at the model grid scale. Their effective values absorb model structure, spatial aggregation, and sub-grid heterogeneity, so estimates from one location do not in general transfer to another.

ADELM treats parameterization as an explicit, learnable component of the land model. Rather than fixing parameter values in advance, each learnable parameter is either a globally shared scalar or an attribute-to-parameter mapping that takes static environmental attributes and returns a spatially distributed parameter field. Because the mapping is differentiable and depends only on attributes that are defined everywhere across the domain, it can be constrained at observation sites and applied at every grid cell — serving the same role as a pedotransfer or attribute-to-parameter function, but with a form that is learned from observations.

This supports two common objectives:

  • Globally shared learning: constrain a single effective parameter value from multi-site observations.

  • Spatially varying learning: learn an attribute-to-parameter mapping that transfers from training sites to unobserved locations.