Methodology overview

FRED's recommendations rest on three published spatial prediction layers. This page explains what they are at a high level, how they compose, and where the underlying research is documented.

#Three spatial prediction layers

FRED organizes its predictive components into three distinct categories. Understanding the taxonomy makes the rest of the docs easier to follow.

  • Learned prediction layer: a machine-learning model trained to predict an outcome from inputs. FRED's harvest cost surrogate is the canonical example. See Harvest cost surrogate.
  • Interpolated empirical layer: a continuous spatial field built from sampled observations using interpolation. FRED's transport circuity model is the example. See Transport circuity model.
  • Optimization layer over external domain data: a decision framework built on top of an external authoritative dataset. FRED's fire-aware Pareto framework wraps the USDA FSim burn probability raster this way. See Fire-aware Pareto framework.

Each layer has different accuracy guarantees, update cadences, and appropriate uses. Treating them as a single category obscures important differences; treating them as the same kind of thing makes for sloppy interpretation.

#Data sources

FRED ingests data from several authoritative sources:

  • C-BREC (California Biomass Residue and Energy Conversion) for cluster polygons, biomass volumes, prescriptions, and terrain class. The cluster catalog is the foundation of every FRED query.
  • FRCS (Forest Residue Cost Simulator) from USDA Forest Service for the harvest cost training data. FRED runs FRCS millions of times during training and queries the surrogate at inference time.
  • OSRM with custom California truck profile for road network routing. Used to build the 445,000-route sample that the IDW circuity model is fit to.
  • USDA FSim 2024 burn probability raster at 30m resolution. Provides the fire-risk input to the Pareto framework.
  • OpenStreetMap California extract for the road network underlying OSRM.

Data provenance is documented per-layer on the following pages.

#How they compose into FRED

A FRED query composes outputs from one or more of the three layers. For a basic cost-only siting query, FRED uses the harvest cost surrogate and the transport circuity layer. For a fire-aware query, the Pareto framework joins the burn probability layer. For a multi-year projection, the same layers run iteratively with cluster depletion.

The composition itself is the contribution of the CS-RAG architecture: dynamic per-query joining of heterogeneous spatial prediction layers, mediated by an agentic reasoning loop.

#Reproducibility

FRED's predictions are deterministic given the underlying model weights and data versions. Two identical queries to the same FRED instance return identical answers; the same query against the same data versions on a different instance returns the same answer modulo numerical precision.

For researchers needing fully reproducible parameterized runs, the prediction layers themselves are available as standalone artifacts:

  • The harvest cost surrogate as a serialized XGBoost model.
  • The IDW transport circuity field as a GeoTIFF raster.
  • The burn probability raster as ingested from USDA (the data provenance is downstream of the original FSim release).

Email contact@biofred.us for access to the artifacts directly, or read the thesis and accompanying papers for the full methodology.