Transport circuity model

An empirical, interpolated spatial layer of road network circuity factors built from 445,000 OSRM truck-profile routes. Produces a continuous statewide raster that converts straight-line distance estimates to realistic chip truck haul distances.

#Overview

Naive transport cost estimation uses Euclidean distance from cluster to facility, often multiplied by a flat cost-per-km. The reality is that chip trucks navigate a constrained road network with grade limits, weight limits, and forest road access that is often missing from consumer-grade routing engines.

FRED's transport circuity model addresses this with an empirical approach: sample real OSRM truck-profile routes across California, compute the circuity factor for each route's origin, then interpolate to a continuous raster using inverse distance weighting (IDW).

#OSRM truck profile

FRED uses a custom Lua profile configuration for OSRM that reflects how chip trucks actually use the road network. The key constraints:

  • Roads weight-restricted below 36 tons are excluded.
  • Road grades above 8 percent are penalized heavily.
  • State highways are preferred for long-haul segments.
  • Seasonal forest roads are penalized to reflect winter restrictions.
  • Bridge load limits are honored.

This profile is documented in the underlying paper and validated against truck operators' real routing decisions. It is materially different from a passenger-car profile; using a car profile would systematically underestimate haul times by 15-30 percent.

#The 445,000 routes

The training corpus for the IDW model is 445,000 OSRM-computed routes across California. Each route was sampled by:

  1. Picking a random origin point in a forested area (sampled from cluster centroids).
  2. Picking up to 10 random destinations representing plausible facility locations.
  3. Computing the OSRM truck-profile route for each (origin, destination) pair.
  4. Recording the resulting circuity factor along with the origin's terrain class, elevation, slope, and county.
CAUTION
An earlier version of the route dataset included duplicates from a treatment-id join error that inflated the count to roughly 7.4 million. The deduplicated count of 445,000 is the figure used in all published results.

#IDW interpolation

Inverse Distance Weighted interpolation is the standard approach for spatial extrapolation when the underlying process has local spatial correlation but no clear parametric form. The IDW prediction at a query point is:

NOTE
CF(q) = Σ (wi · CFi) / Σ wi, where wi = 1 / d(q, samplei)p

FRED uses p = 2 and the 12 nearest sampled routes for each prediction, chosen by cross-validation. The output is rasterized to a 1km grid across California and also stored per- cluster on the materialized supply view.

#Why origin dominates

A key methodological finding from FRED's transport research: origin location explains 87 percent of the variance in CF. The destination effect, conditional on origin, is much smaller.

This is intuitive once you see it. The road quality at the origin determines how much switchback navigation is required to reach any nearby highway. Once on the highway, the route to a destination is mostly straight. So all routes from a given origin share most of their circuity, and the destination just adds a small variance term.

The practical consequence: FRED can pre-compute a single CF value per cluster (sampled at the cluster centroid) rather than a circuity matrix between every origin-destination pair. This is the difference between roughly 2 million stored values and roughly 2 million squared, which is computationally impossible.

#Limitations

  • Sample density variation. Some remote terrain has fewer sampled routes, so the IDW prediction in those areas has wider uncertainty. FRED's responses flag this when relevant.
  • OSRM truck profile is California-tuned. The weight, grade, and seasonal constraints reflect California regulations. Applying the model to other states would require re-sampling routes with the relevant local truck profile.
  • Static road network. The OSRM data is a snapshot. Construction projects, washed-out bridges, and new forest road closures are not captured until the next refresh.
  • Truck-specific time and fuel constraints not modeled. The circuity factor captures the road network geometry. Driver hours-of-service rules, fuel stops, and other operational constraints are factored into the cost layer separately.