# PERSEUS libcbm vs GCBM cross-state intercomparison (CONUS, 2026-06-02)

PERSEUS multi-model forest-carbon intercomparison dataset: 48 lower-48 US
states under the canonical B1.1 v6 and FIA-EXPNS-anchored B1.3
inventory-stratification hypotheses, plus 6-state libcbm vs GCBM gap
matrices, per-pool decompositions, F3 Q10 mean-annual-temperature
sensitivity sweeps, and the five publication figures supporting the
inventory-stratification methods finding.

This deposit is the companion dataset to the PERSEUS methods note
"Inventory stratification dominates the libcbm-vs-GCBM engine gap" at
https://holoros.github.io/perseus-forest-intelligence/methods/inventory-stratification/
and is intended to be cited from the v1.3 manuscript in preparation
(Weiskittel et al., in prep).

## What is here

```
perseus_libcbm_vs_gcbm_v1/
├── README.md                                    this file
├── CITATION.cff                                 machine-readable citation
├── data_dictionary.csv                          column descriptions for every CSV
├── zenodo_metadata.json                         Zenodo API metadata
├── figures/                                     5 publication figures (PDF + PNG)
├── matrices/                                    7 ratio + decomposition matrices
├── libcbm_pools_b13_fia/                        48 per-state libcbm year-5 pool CSVs
└── sweeps/                                      2 F3 Q10 sensitivity sweep CSVs
```

## Headline findings

**1. The +24% libcbm-vs-GCBM engine gap is reproducible across 5 of 6 states
with full GCBM aggregates.** Under the canonical B1.1 v6 parity (uniform-FT
inventory + Boudewyn vol-to-biomass + LCMS disturbance events), WA, MN, IN,
ME, and OR all sit at libcbm/GCBM density ratios between 0.74 and 0.80
(gaps of 20.5 to 25.6 percent). Georgia is the lone outlier at ratio 1.05,
consistent with its warm-donor Stage 1 placeholder AIDB.

**2. Inventory stratification is the dominant component of cross-engine
uncertainty.** Replacing the uniform-FT B1.1 inventory with B1.3 FIA EXPNS
(the canonical FIA Total Area Estimator) at n=39 of CONUS states adds +6,760
TgC to the CONUS-wide carbon stock — a +14 percent shift on 246.5 Mha. Most
of what was attributed to engine differences was an inventory stratification
choice.

**3. The +24% engine gap carries a regional DOM-pool fingerprint.** The
Pacific Northwest (mean slow-soil pool 184 Mg/ha across WA, OR, CA) and the
cool-moist Atlantic Maritime Northeast (mean 143, range 113-208) carry the
highest libcbm slow-soil carbon. South and Lake States sit around 105 Mg/ha.
Top-quartile slow-soil states are exactly the cool-moist climates: VT, OR,
NY, NH, WA, CA, ID, ME. An F3 Q10 mean-annual-temperature sensitivity
sweep on Oregon (MAT 4 to 13 C) confirms the slow-soil overshoot is
spinup-driven, not runtime Q10-driven: even at MAT 13 C, OR libcbm/GCBM
stays at 1.35.

## Method note (short)

48-state libcbm pool outputs were computed with the GCBM2hpc pipeline at
`github.com/holoros/GCBM2hpc` (head `1362169` at time of this deposit). For
every state we (a) downloaded the FIA TREE + COND + POP_STRATUM panel via
rFIA, (b) fit Boudewyn vol-to-biomass coefficients and component proportions
from the state FIA, (c) cloned an AIDB donor with Q10 F3 correction, (d)
computed FIA EXPNS expansion-factor areas per FT group, (e) built the GCBM
input db and the libcbm SIT bundle under B1.3 FIA EXPNS inventory, and (f)
ran libcbm 5-year baseline. The 6-state GCBM aggregates were produced
spatially on OSC Cardinal via the moja FLINT containerized GCBM at
1-degree WGS84 tiles.

Detailed reproduction recipe in the GCBM2hpc README.

## Citation

If you use this dataset, please cite both the deposit DOI and the methods
paper:

> Weiskittel, A. R., et al. (in prep). Inventory stratification dominates
> the libcbm vs GCBM engine gap in CBM-CFS3 family forest carbon
> intercomparisons across CONUS. Manuscript in preparation.

> Weiskittel, A. R. (2026). PERSEUS libcbm vs GCBM cross-state
> intercomparison (CONUS) [Dataset]. Zenodo.
> https://doi.org/10.5281/zenodo.XXXXXXX

Pipeline source: github.com/holoros/GCBM2hpc.

## License

CC-BY 4.0 (Creative Commons Attribution 4.0 International).

## Provenance

Compute on the Ohio Supercomputer Center Cardinal cluster (allocation
PUOM0008). Engines compared: CBM-CFS3 (Canadian Forest Service), libcbm
(CFS Python/C++ reimplementation), and GCBM on moja FLINT (spatially
explicit per-pixel). Auxiliary CONUS data products: TreeMap 2022 (USFS),
LCMS disturbance v2024 (USFS), FIA DataMart, NOAA 1991-2020 climate
normals, EPA Level III ecoregions.