Where the width comes from — and what to contribute. The interval is dominated by what published studies fail to report. Our measured uncertainty budget (σ as % of strength when a factor goes unreported), each with the peer-reviewed DoE that would collapse it:
💧 Filament water state (moisture, drying, storage age, ambient RH) — ~14%, the biggest uncontrolled term. Neat-filament sensitivity is now measured for 12 materials; the missing piece is a printed-coupon dry-vs-wet DoE: one PLA + one PA, dried / week-ambient / saturated, ≥3 replicates, conditioning stated. (~12 coupons, one print day.)
🧱 Trans-layer (Z) scatter — 11%: upright coupon series, ≥3 reps — especially any family beyond PLA/ABS (our Z data covers only those two).
🛏 Plate layout & interlayer time — 10%: the same coupon printed solo vs ×4 vs ×9 per plate, 3 reps. Nobody publishes this; ~9 coupons.
📐 Geometry: specimen standard (7%) & thickness (8%) — kept as separate factors: one material across ISO 527-1B vs ASTM D638-I and two thicknesses. (D638 reads +13–20% higher on the same material.)
🧭 Toolpath / G-code detail — ~5%: seam position, per-layer raster order, crosshatch alternation — the path semantics papers leave unstated (our forensic pass had to correct several at ±5–8% effect). Contribute by publishing the G-code or full toolpath parameters with any DoE.
🌀 Part-cooling fan on hot-bonding families — a first-order lever (this model: ABS collapses ~58% at 100% fan) calibrated from ONE study in which fan, nozzle and bed moved together — so treat it as an upper bound. Wanted: an independent fan sweep — one hot family (ABS/ASA/PC), fan 0/50/100%, everything else fixed, ≥3 reps ≈ 9 coupons. Cheapest high-value DoE on this list.
🌡 Machine: temp set-vs-actual (5%) and true speed (5%) — separate factors with separate fixes: thermocouple nozzle checks, and motion logs on accel-limited short coupons — nearly free, rarely done.
🎯 Replicate floor: 3–4.5% in-plane — the irreducible target. Fully specified studies already land ON it in our held-out tests; every field you report moves your data toward that floor.
🌗 Inside the grey — the known unknowns. Downside (no data yet): pigment & additive packages within a "family" · filament diameter tolerance · nozzle wear (line-width drift) · storage aging · printer motion quality · test-lab systematics. Upside — factors that make prints BETTER than nominal: annealing / crystallinity from warm chambers & slow cooling · above-spec resin batches (we measured one filament +9.4% over its own datasheet) · truly-dry filament · hot, short-interlayer-time welds on small fast parts. Each of these is a dataset away from leaving the grey.
🌑 Non-attributed ~25%: what remains after every named source — resin batch, additive packages within a "family", lab & machine systematics. Sized by closure: observed cross-study residual 33.6% minus the named budget (RSS 22.5%). Only shrinks via round-robins (same spool, many labs/printers) — the hardest and most valuable contribution of all.
Complete metadata is the entire game: report infill %, pattern, walls, flow, measured part density, moisture/drying, fan, temps and your data enters the model at floor accuracy. How to contribute →
Factor catalogue — where the digital twin stands. The input schema is Open3DPP (99 columns, Open3DCP-derived): one common dataset format for material × process × geometry across all FDM systems. Factors graduate from recorded to modeled as open correlating data lands.
Modeled today (moves the number): material family · brand/product anchors (API & CLI) · infill density · perimeter walls + top/bottom skins · layer height · raster angle · build orientation (flat / upright-Z) · nozzle temperature · flow ratio · part-cooling fan (hot-bonding families; single-source calibration) · print speed · measured part density (supersedes inference when reported).
Recorded in the schema, awaiting correlating data: infill pattern (first stated-density dataset landed: gyroid strongest on PETG @50%; a second family graduates it to modeled) · bed temperature · printer system/model · filament moisture & drying state (per-material sensitivity measured; not yet a prediction input) · part geometry beyond the reference coupon.
Planned (geometry & structure blocks): nozzle diameter / line width · interlayer time & thermal mass · chamber / ambient humidity · filament batch & age · void fraction, degree-of-healing, crystallinity (ICME mediators).
How this works. A physics anchor (material identity × load-bearing fraction from infill/walls/flow × orientation × thermal bonding, Gibson–Ashby cellular scaling) corrected by a ridge model trained on 173 published measured coupons, with conformal 90% intervals. Intervals are infill-conditional: sparse parts genuinely scatter more, so their bands are honestly wider. Predictions are referenced to a standard 13 × 3.2 mm tensile coupon (ASTM D638 Type I) with slicer-default shell (selected walls + 0.8 mm solid top/bottom skins; walls = 0 means a bare laminate) — real part geometry shifts wall share.
Walls enter the load model (0.48 mm line each); infill pattern is not yet modeled — no open dataset states pattern AND density together. If you have such data, we want it.
Honest limits. Cross-study error on held-out studies is 1–13 MPa depending on how completely the process is specified; the interval, not the point, is the product. Model + data: open source, predictions CC BY 4.0.
Disclaimer. Estimates for orientation only — research preview, provided AS-IS without warranty of any kind. NOT a substitute for testing. Do not use for safety-relevant, load-critical, or regulated design. See Terms.