XYZ BIMCollective · Portugal Research
xyzbim.eu / pesquisa / carbono-bases · study · 2026-09-03

The same material is worth depending on the database — but the databases are not to blame

Comparative study across three free embodied-carbon databases — ICE (UK), KBOB (Switzerland) and ÖKOBAUDAT (Germany) — pairwise, per kg, modules A1–A3, replicable by scripts. Conclusion in one sentence: for the same manufacturing route the databases converge (−6%); for different routes they diverge 2.4× within the same database. Choosing a carbon factor is choosing a supply chain. It grew out of the digital product passport dossier.

databases compared3 ICE · KBOB · ÖKOBAUDAT
EAF steel, ICE vs KBOB−6% same route, independent databases · measured
UK vs EAF steel, within ICE2.4× different routes, same database · measured
pairs analysed6+2 6 ICE×KBOB + 2 OBD triangulations
supporting literature10 papers with DOI (2011–2025)
Verdict — the databases agree; it is the products that differ. For the pair most sensitive to the manufacturing route (rebar steel), two independent databases (ICE Educational and KBOB) give the same value within ±6% when the route is the same (electric arc furnace with scrap, European) — and the German generic from ÖKOBAUDAT (1.57 kg CO₂e/kg) sits exactly between EAF (0.73–0.77) and the British blast-furnace average (1.72), where a mixed route should be. The variation that scares people in carbon numbers is not “database error”: it is geography and process. The practical consequence — for the digital product passport and for whoever specifies — is that every value needs traceability: source, version, unit and route.

1 · Question and hypotheses

Design: controlled, pairwise comparison, with inclusion criteria fixed before looking at the results.

Question: when two people calculate the embodied carbon (A1–A3) of the same material with different public databases, does the difference measure database error or real supply-chain differences?

H1 (intra-route convergence): same material + same manufacturing route + comparable geography → independent databases converge (±15% [inferred]). · H2 (inter-route divergence): different routes → the variation within the same database exceeds the variation between databases for the same route. · H0: the differences between databases are random with respect to the route.

2 · Methodology

Sources with registered versions, programmatic extraction (scripts at the end), normalisation per kg, exclusions logged instead of hidden.

databaseversionaccesslicence
ICE Database (Circular Ecology, UK)Educational V5.0 (Jun 2026)xlsxeducational (expires 30 June 2028) — comparative study
KBOB (Swiss government)Ökobilanzdaten im Baubereich V9.0official xlsxpublic
ÖKOBAUDAT (German government)REST API, general datastock, queried 3 September 2026JSONpublic

Inclusion criteria (pre-registered): GWP-total indicator, modules A1–A3 (summing A1+A2+A3 when published separately); normalisation to kg CO₂e/kg; pairs with an ambiguous declared unit are excluded and logged — we excluded OBD's generic “Facing brick” (555.2 kg CO₂e with an unidentifiable implicit unit) and all manufacturer products declared per piece or per m². Documented exception: concrete — KBOB declares per kg, OBD per implicit m³; converted with the KBOB density (2,300 kg/m³), declared as an assumption.

3 · Results

Left: ICE × KBOB (both per kg). Right: ÖKOBAUDAT triangulation on the two anchor materials.

material (kg CO₂e/kg, A1–A3)ICEKBOBΔcomparable route?
Structural concrete (general)0.1030.089+17%yes (UK/CH national averages)
Rebar steel — recycled EAF0.730.773−6%yes (EAF Europe)
Rebar steel — UK average (blast furnace)1.720.773 (EAF/CH)+123%no
Softwood solid timber (no carbon storage)0.2630.134+96%partial (drying/process)
Glass wool1.531.04+47%partial
Common ceramic brick0.2130.254−16%yes
ÖKOBAUDAT triangulation (generics)OBDreading
Rebar B500B (A1+A2+A3)1.574 /kgbetween EAF (0.73–0.77) and the UK average (1.72) — intermediate position consistent with a mixed German/European route
Concrete C30/37 (A1–A3)257.7 /m³vs KBOB 0.089 × 2,300 = 204 /m³ → +26%

Reading. H1 confirmed in the critical case: intra-route dispersion (±6%) is an order of magnitude smaller than inter-route (2.2×) [measured]. H2 confirmed: within ICE itself, UK vs EAF rebar differs by 2.4× — more than any between-database difference for the same route. H0 rejected for steel. Concrete (+17/+26%) and brick (−16%) — heavy materials, national averages — converge reasonably. Timber (+96%) and glass wool (+47%) — where biogenic accounting, drying and system boundaries vary most — diverge for procedural, not spatial, variables.

4 · What the literature already knew (and what this study adds)

Ten papers with DOI position the result — variation between databases is a documented problem; isolating the route effect is this study's addition.

“Reported results display large unjustifiable variations”; the variables framework includes “spatial disparities” and “procedural inconsistencies” — 244 normalised case studies. “A systematic investigation into the methodological variables of embodied carbon assessment of buildings”, Renewable and Sustainable Energy Reviews, 2021 · doi.org/10.1016/j.rser.2021.110840
“LCA method and for instance geographical location are incompatibilities also revealed in embodied energy and embodied carbon assessments.” “Embodied energy and embodied carbon of structural building materials: Worldwide progress and barriers”, Energy and Buildings, 2020 · doi.org/10.1016/j.enbuild.2020.110612

Completing the picture: the classic generic↔specific reference (Int. J. LCA, 2015), the 401-citation review on the lack of consensus in practice (Energy & Buildings, 2017), the ongoing international harmonisation (IEA EBC Annex 72), EPD comparability (J. Cleaner Prod., 2011) and the existing reviews of data (2012) and of insulation (2021). This study's addition is the controlled demonstration that, in the material most sensitive to the route, the between-database variation for the same route is smaller than the between-route variation within the same database — replicable by scripts, with the three free databases.

5 · Limitations (all of them)

  1. Small n (6 pairs + 2 triangulations), curated — not systematic; the selection may introduce bias.
  2. Point-in-time versions (Jun 2026 / V9.0 / single-day query) — no time dimension.
  3. ICE Educational: educational licence; these values serve a comparative study — commercial data work uses only unrestricted public sources (KBOB, ÖKOBAUDAT, EPDs).
  4. Implicit units in OBD for generics — conversions declared as assumptions.
  5. Non-harmonised biogenic accounting (timber) — likely cause of the +96% [inferred].
  6. Single author; automated extraction without independent verification (the public scripts partly compensate).

6 · Conclusion and practical implication

Three conclusions. (1) For the same product and route, the public databases agree — they can be trusted provided source, version, route and unit are declared. (2) The dominant variation lies in route and geography, not in database quality: choosing a carbon factor is choosing a supply chain. (3) For the digital product passport, the operational requirement is not “one official database” but traceability per property — that is why the connector epd_lookup.py (published with this study) emits each value with source, UUID, version, unit and module.

Tools from this study, to use now: passport readiness meter (load an IFC, see what is missing) · digital product passport dossier (laws, standards, timeline, 893 papers)

Replication: the scripts that extracted every number are published with the study — parse_ice.py (ICE × KBOB; requires the ICE Educational V5.0 xlsx, obtained by registering with Circular Ecology, and the public KBOB V9.0 xlsx) and epd_lookup.py (queries the open ÖKOBAUDAT API and emits the IFC pset with UUID and version). Python 3 with openpyxl; each run leaves its audit JSONs next to the data — this study's pass is preserved in four lookup_*.json files with the full API response. Note: epd_lookup.py disables TLS certificate verification to get through corporate proxies; on a normal network it may return the SSL context by default. Nothing in this study is a second-hand number — run the scripts and compare.