Preprint WI-PRE-2026-01 · rev. 1.0 · 2026-08-01 · not peer reviewed
An open, calibrated screening model for exterior timber service life
Full manuscript for Wood Intelligence service-life screening method (WI-SL v1.2.0). It states the formulation, the uncertainty treatment, the calibration procedure and the blind validation result in the form a reviewer needs to check them. Everything cited here is published on this site as data, equations or code output.
Status: preprint. This manuscript has not been peer reviewed and carries no DOI. It is published in this form so that the method can be criticised before it is used, not after.
Abstract
What was done and what came out
Specifying exterior timber requires an estimate of how long a given species will remain serviceable at a given site. Existing guidance is either qualitative (EN 350 durability classes) or proprietary. We present WI-SL v1.2.0, an open factor-method model in the ISO 15686-8 tradition that maps climate reanalysis inputs, façade orientation, macro-exposure and an ordinal species resistance profile onto a service-life estimate for 12 timber species and modified-wood products.
Input uncertainty is propagated by Monte Carlo simulation (1200 draws, Gaussian, truncated at ±3σ, deterministically seeded), so every output is reported as P10/P50/P90 rather than a single number. The uncalibrated formulation was systematically optimistic on external observations: bias +19.6 yr, MAE 22.3 yr, P10–P90 coverage 0% against an 80% target (WI-VAL v1.1.0, n = 24).
We therefore add an explicit, published correction layer (WI-CAL v1.1.0): a log-space multiplicative factor per Köppen main class, shrunk towards a global factor with 2 pseudo-observations, plus a ×4.5 inflation of the Monte Carlo residual σ. Under blind leave-one-out refitting the corrected model achieves bias 0 yr, MAE 9.5 yr, RMSE 12.6 yr, coverage 83.3% and r = 0.56. The correction removes the level error; it does not improve case-by-case ordering, which remains the principal open problem and is limited by the size of the public field-observation record, not by the formulation.
1
Introduction
EN 350 classifies natural durability against biological agents on an ordinal 1–5 scale, and EN 335 assigns use classes by exposure. Neither yields a service life in years for a specific façade at a specific coordinate, which is the quantity a specification actually needs. ISO 15686-8 provides the factor-method structure for converting a reference service life into an estimated one, but leaves the factors to the practitioner.
This work fills that gap with a fully published parameter set: 15 equations, 14 indexed standard clauses, and every constant readable at /method. The design constraint was reproducibility before accuracy: a third party must be able to recompute any published number without access to our codebase.
2
Method
The model evaluates 15 equations in fixed order: climate load terms (UV, driving rain, humidity, chloride, freeze–thaw) derived from ERA5 reanalysis and Köppen classification; orientation and macro-exposure modifiers; an ordinal species resistance profile derived from the EN 350 durability class and known extractive chemistry; a reference service life per durability class; and the factor-method combination that yields the deterministic estimate.
Uncertainty enters as a declared standard deviation on each input group. The Monte Carlo pass draws 1200 realisations (mulberry32 PRNG seeded from an FNV-1a hash of the input brief), so a given brief always returns identical percentiles. A one-at-a-time sensitivity pass of 400 draws per group attributes output spread to individual parameters and is shown next to every result.
| ID | Equation | Basis |
|---|---|---|
| E1 | UV load | Orientation weighting after EN 927-1 exposure severity; climate term from ERA5 reanalysis. |
| E2 | Moisture load | Driving-rain exposure per EN ISO 15927-3; wetting/drying logic per EN 335 use class 3. |
| E3 | Chloride and freeze-thaw load | ERA5 2 m temperature series; distance-to-coast chloride gradient. |
| E4 | Biological load | Decay-hazard logic after EN 335 and Scheffer-type climate indices. |
| E5 | Weighted performance | Wood Intelligence formulation. Ordinal, not a physical damage model. |
| E6 | Service life | Factor-method structure after ISO 15686-8, with a single aggregated performance factor. |
| E7 | Maintenance interval | Recoat intervals after EN 927-1 thin-film system classes. |
| E8 | Silvering | Photodegradation rate as a function of cumulative UV dose. |
| E9 | Confidence | Wood Intelligence uncertainty heuristic. Not a statistical confidence interval. |
| E10 | Heuristic service-life range | Deterministic spread. Superseded by the Monte Carlo percentiles in E11–E13. |
| E11 | Input sampling | Standard Monte Carlo propagation, GUM Supplement 1 (JCGM 101) structure. |
| E12 | Output percentiles | Non-parametric quantiles; no distributional assumption on the output. |
| E13 | Per-parameter uncertainty share | One-at-a-time variance decomposition. An approximation of a Sobol first-order index. |
| E14 | Empirical calibration | Fitted against WI-VAL v1.0.0. Applied to service life only, as a published layer on top of E1–E8, so the physics and the correction can be audited separately. |
| E15 | Residual σ inflation | Chosen so that leave-one-out P10–P90 coverage on the benchmark reaches the 80 % target. Declared input uncertainty alone understated real scatter roughly fourfold. |
Table 1 — equation register. Full expressions and term definitions at /method.
3
Data and benchmark construction
The benchmark WI-VAL v1.1.0 (2026-08-01) contains 24 external observations of exterior timber reaching the end of serviceable condition. Observation types: field-trial, in-service-survey, standardised-test, service-life-database. Source types: institute-report, peer-reviewed, database. No case is produced by the model it tests.
Two mapping steps introduce error that we retain rather than hide. Where a source reports a range, the midpoint is used and the range is recorded. Site climate is mapped to the nearest Köppen-equivalent reference climate in the platform index, not to the exact station, and that mapping error is included in every reported residual. The sample is small and skewed towards northern Europe; aggregate metrics should be read alongside the per-zone table.
4
Results
| Model state | n | Bias (yr) | MAE (yr) | RMSE (yr) | MAPE | Coverage | r |
|---|---|---|---|---|---|---|---|
| Uncalibrated (physics only) | 24 | +19.6 | 22.3 | 26.5 | 121.4% | 0% | 0.36 |
| Blind — leave-one-out calibration | 24 | 0 | 9.5 | 12.6 | 37.5% | 83.3% | 0.56 |
| In-sample calibrated | 24 | -0.5 | 8.5 | 11.2 | 33.1% | 87.5% | 0.63 |
Table 2 — aggregate accuracy. The middle row is the figure to judge the model by: the correction is refitted with each case removed before that case is predicted. The bottom row is in-sample and flatters the model by construction.
| Köppen zone | n | Bias | MAE | RMSE | MAPE | Coverage | Factor |
|---|---|---|---|---|---|---|---|
| A — tropical | 4 | +2.6 | 3.5 | 4.7 | 37.1% | 75% | ×0.3268 |
| B — arid | 5 | +0.9 | 3.8 | 5.4 | 21.5% | 100% | ×0.441 |
| C — temperate | 9 | -0.6 | 12.1 | 13.4 | 41.7% | 88.9% | ×0.7014 |
| D — continental | 6 | -3.6 | 10.5 | 14 | 27% | 83.3% | ×0.6906 |
Table 3 — accuracy by Köppen main class, with the fitted correction factor. 5 zones carry a factor; unfitted zones fall back to the global factor ×0.547. Rows with n < 4 are reported for transparency, not for confidence.
| Species | n | Bias | MAE | MAPE | Coverage | r |
|---|---|---|---|---|---|---|
| Thermally Modified Ash | 5 | +15 | 15 | 72.5% | 100% | 0.98 |
| European Oak | 3 | -8.1 | 8.1 | 15.6% | 100% | 0.80 |
| Siberian Larch | 4 | -6.7 | 8.4 | 30.9% | 75% | 0.99 |
| Norway Spruce | 4 | -6.4 | 7.5 | 33.9% | 50% | 0.77 |
| Acetylated Radiata Pine | 4 | -2.9 | 2.9 | 8.2% | 100% | 0.98 |
| Western Red Cedar | 4 | +0.2 | 7.6 | 23.1% | 100% | 0.19 |
Table 4 — residuals by species. Species-level bias that survives climate-zone correction indicates a resistance-profile error rather than a climate error, and marks where the next parameter revision should be spent.
Two findings are robust to how the tables are read. First, the physics-only formulation was optimistic by roughly a factor of two, and a flat multiplicative correction removes almost all of that bias without altering a single equation — the error was in the level, not in the ranking mechanism. Second, reaching honest 80% interval coverage required inflating the residual σ by ×4.5, meaning the declared input uncertainties alone understated real scatter roughly fourfold.
Correlation remains modest (r = 0.56 blind). The correction fixes the average, not the case-by-case ordering. Closing that gap requires more field records with documented detailing, not additional mathematics.
5
Limitations
- Service life is empirically calibrated (E14) against a 24-case benchmark. The correction removes the average bias; it does not make the model a validated damage model, and the zone factors for tropical, arid and polar climates rest on two cases or fewer.
- The calibration is fitted on the same benchmark used to report accuracy. Leave-one-out metrics are published on /validation and are the correct figures to judge the model by.
- Resistance profiles are expert-assigned ordinal scores, not laboratory measurements. Only the EN 350 durability class behind them is a standardised value.
- The confidence band (E9) is a rule-based heuristic, not a statistical interval. The P10/P50/P90 figures come from Monte Carlo propagation (E11–E13) of declared input uncertainty — they are as good as those declared σ values, which are modelling assumptions until replaced by measured field variance.
- Input uncertainties are treated as independent. Real correlation between rain, humidity and biological load will make the true spread wider than modelled at the tails.
- Detailing quality — ventilated cavity, end-grain sealing, fall, ground clearance — dominates real outcomes and is only partially represented.
- Treatment performance assumes a correctly applied system maintained to its stated interval. Product-specific data is not used.
- Climate inputs are zone-level, not site-level. Local topography, shading and coastal microclimate are not resolved.
6
Data, code and reproducibility statement
Every number in this manuscript is regenerated from the published artefacts below. The model is deterministic: identical inputs return identical percentiles, because the pseudo-random stream is seeded from a hash of the input brief.
- /method — full equation set, all constants, declared input uncertainties (WI-SL v1.2.0, released 2026-08-01).
- /validation — the benchmark, per-case residuals and a downloadable JSON validation record.
- /wood-index — the precomputed reference dataset (CSV and JSON) used for cross-species comparison.
- /specify — any individual result can be exported as a calculation record embedding the full method spec and the intermediate trace.
Licence: CC BY 4.0. Reuse, redistribution and independent recalculation are permitted with attribution.
7
Funding and competing interests
Wood Intelligence was initiated and is funded by SiOO:X Wood Protection. The funder has no influence on model parameters, benchmark selection or published results, and no product-specific data is used anywhere in the method. Treatment effects are modelled as a generic, correctly applied and maintained protective system, not as a named product.
This funding relationship is a competing interest and is declared here for that reason. The mitigation is procedural rather than rhetorical: the full parameter set, the benchmark and every residual are public, so any bias in favour of treated timber would be visible in Table 4.
8
References
- EN 350:2016 — Durability of wood and wood-based products — Testing and classification of durability to biological agents. Used for: Durability class → reference service life (E6)
- EN 335:2013 — Durability of wood and wood-based products — Use classes. Used for: Use class assignment and biological load (E4)
- EN 927-1:2013 — Paints and varnishes — Coating materials for exterior wood — Classification and selection. Used for: Exposure severity and maintenance interval (E1, E7)
- EN ISO 15927-3 — Hygrothermal performance of buildings — Driving rain index. Used for: Moisture load (E2)
- ISO 15686-8 — Buildings and constructed assets — Service life planning: reference service life and service-life estimation. Used for: Factor-method structure (E6)
- ERA5 — Copernicus Climate Change Service, ERA5 hourly reanalysis on single levels. Used for: Temperature, humidity, precipitation and irradiance inputs (E1–E4)
- Köppen-Geiger — Beck et al. (2018), Present and future Köppen-Geiger climate classification maps. Used for: Climate zone assignment
Benchmark sources are listed per case, with reference and link where public, on /validation.
Cite
How to cite this preprint
Wood Intelligence (2026). An open, calibrated screening model for exterior timber service life (WI-SL v1.2.0). Preprint WI-PRE-2026-01, rev. 1.0. Wood Intelligence. https://woodintel.io/paper
@techreport{woodintelligence_wi_pre_2026_01,
title = {An open, calibrated screening model for exterior timber service life},
author = {{Wood Intelligence}},
institution = {Wood Intelligence},
type = {Preprint},
number = {WI-PRE-2026-01},
version = {1.0},
year = {2026},
month = {8},
url = {https://woodintel.io/paper},
note = {Method WI-SL v1.2.0; benchmark WI-VAL v1.1.0; licence CC BY 4.0. Not peer reviewed.}
}