Where do critical raw materials really come from?

For each material the tool answers three questions with real data: where is it mined, where is it refined, and who actually trades it with whom — globally, in both directions. Pick any material and any country. The European Union is available as one option among others — an aggregation you can look at, with a dedicated correction built for it (below) — but the tool is not EU-specific.

The layers

• Reserves / potential — world reserve shares by country, from USGS Mineral Commodity Summaries (approximate): where a material could come from, exploited or not — it lights up under-developed holders such as Bolivia (lithium), Vietnam and Guinea. Conservative, though: USGS reserves are proven and economic, so frontier deposits still classed as "resources" (e.g. Greenland's rare earths) stay dim, and byproduct or abundant materials (gallium, germanium, hafnium, silicon, arsenic, and a few others USGS leaves blank) carry no figure.
• Mined — world mine-production shares by country, from USGS Mineral Commodity Summaries (approximate).
• Refined / processed — where the ore is turned into usable metal. Sourced best-per-material: for the metals it reports, BGS World Mineral Statistics gives annual refined/smelter production by country (copper, cobalt, nickel, alumina, magnesium, germanium, arsenic — 2019–2024, with a year slider); otherwise the IEA Critical Minerals Outlook 2026, EU CRM 2023 (JRC/SCRREEN2), or USGS, whichever is the strongest published figure. Each profile shows the actual source it used.
• Traded — complete bilateral trade between every country pair, both directions, for each year 2018–2024 (selectable with the year slider): primary source UN Comtrade, in its reconciled, mirror-harmonised form CEPII BACI. The map, globe and flow view draw this directly; arrows are coloured by what is shipped — raw ore vs refined product. 2024 is the newest complete year (BACI release V202601, Jan 2026). The slider also offers a provisional 2025 and a directional 2026 built from our own Comtrade reconciliation for the years BACI hasn't released (see Validation & failure cases below). The mine and refine layers are undated reference, not yearly.

Why the refiner is not the source

Under customs rules, refining counts as "origin": when a country imports an ore, refines it and re-exports, the metal takes the refiner's nationality and the trade record stops there. One layer deeper, the picture changes — the apparent origin is often just the chokepoint, not the mine:

So an apparent China dependence on cobalt is, upstream, a Congo mining dependence funnelled through a Chinese refinery — China is the chokepoint, not the source. This mine-vs-refiner gap is well established in IEA and USGS supply-chain work; what this tool adds is to overlay all three layers — mine, refinery and customs trade — in one material-level interface, where mainstream dashboards (OEC, Resource Trade Earth, Comtrade) typically expose just one. It draws the implied mine → refiner → buyer chain, while being explicit (see Limitations) about what is measured versus referenced.

A method note: correcting the EU's import statistics

The same materials drive a classic correction specific to the European Union — the analysis this project grew out of, documented here (the live tool itself is global and country-agnostic). For extra-EU imports, the Eurostat Comext partner field is the country of origin. That single fact separates two views of the same data:

• Naive — by importing member state. Ranks dependence by which EU country clears the goods. Distorted by the Rotterdam/Antwerp effect: imports are customs-cleared in NL/BE ports (or refined in hubs such as Umicore in Belgium) but bound elsewhere. This measures logistics geography, not dependence. It is the trap.
• Corrected — EU as one entity, by country of origin. Aggregates the 27 member states into a single importer and ranks by true origin. This is the answer.

Across the thirty-two materials the naive and corrected panels disagree on the single most important fact — who the bloc depends on — and only the corrected one is true. China is the real origin for ten, far from a majority; the two sharpest dependencies are not China at all (beryllium is 100% United States, boron 98% Turkey), and the rest span South Africa, Brazil, Japan, Vietnam, Guinea, Gabon, Russia, Chile, Algeria/Qatar, Norway, Mexico, Kazakhstan and Tajikistan — almost none visible in the member-state view. A sample:

Why the EU correction can't be faked from the raw download

Because some codes (gallium, germanium) carry 15 years of data, dependency is a trajectory, not a snapshot: gallium's China-origin share runs 96.8% (2022) → 85% (2023) → 68% (2024) as Canada and Russia step in — China's July-2023 export controls visible directly in the customs record.

Limitations — what is measured vs referenced

This is an overlay of three different measures, not one observed supply chain. Read each row as three lenses laid over each other, not a single flow:

• Three concepts, three vintages. Mine production (USGS, physical, approximate), refining (IEA, processing, approximate) and trade (BACI, customs value, 2018–2024) measure different things; the EU lens uses Comext 2024. A row juxtaposes them — it does not reconcile them into one measured chain.
• HS6 resolution. Gallium and germanium share the six-digit code 811292 (both report under it), so their trade flows are identical and must be treated as one line, not two. 811292 is officially a catch-all also covering hafnium, indium, niobium, rhenium and vanadium; this atlas resolves hafnium via its distinct EU CN8 line (8112 31 00), so hafnium is a separate series here — with the caveat that non-EU hafnium trade reported only at HS6 is buried in 811292 and thus undercounted. Bottom line: gallium ≡ germanium in trade; hafnium is separate but its trade figure is a lower bound.
• Re-exports. Customs trade cannot separate a true refiner from a re-export hub. Hong Kong, Singapore, the Netherlands and Switzerland can surface as top "exporters/refiners" of a material they merely trans-ship. The map's role colours are trade exposure, not proof of physical processing.
• Raw vs refined HS codes. A material's ore and its refined form sit under different HS codes. The map derives miners from the ore code where it is traded, and falls back to USGS/IEA reference shares where it is not — so the data source behind a role can change from one material to the next.
• The flow chart's upstream links are structural. mine → refiner links are drawn from reference shares, not observed flows; only refiner → buyer links are real trade.
• Value, not tonnage. BACI is in US dollars; the top exporter by value can differ from the top by quantity (price and grade effects). Value-vs-tonnage divergences are shown per material in the detail panel.
• Colour intensity is sqrt(volume) for visibility, so a dominant supplier reads as strong but not overwhelmingly so; treat the choropleth as ordinal, not a precise linear scale.
• The satellite layer names where, not which. The Maus mine polygons are all-commodity — no per-mineral label — so the footprint is a physical cross-check, never a material-level source. We quantified how far open data can close that gap: overlaying a single peer-reviewed mine-facility database (Jasansky et al. 2023) labels only ~17% of the footprint and barely 4% ties to a tracked critical material — but stacking thirteen open registers lifts that to 63%, and grouping neighbouring polygons into mining districts reaches ~73%, matching the published state of the art (Maus et al. 2026 / Mine the Gap). What does not improve is the per-mineral resolution: even fully labelled, the open taxonomy resolves ~11 broad commodity classes and rarely names lithium, cobalt, rare earths or tungsten as a primary product, and a census of the unlabelled remainder shows it is ~81% coal and construction minerals, not hidden criticals. This is why material identity here comes from USGS/IEA production shares and trade, not imagery.

Provisional 2025 — a self-built nowcast

BACI lags ~1.5 years, so for 2025 (flagged 2025* on the slider) the atlas does not wait for CEPII — it uses a nowcast built here from raw UN Comtrade. A small Python pipeline (reconcile/) replicates BACI's method: it pulls both mirror reports of every flow (exporter FOB, importer CIF), estimates CIF/FOB markups, weights each reporter by a variance-components reliability model (E[discrepancy²] = var_i + var_j), and reconciles by inverse-variance averaging on logs. Validated against CEPII's official 2024 on what this atlas shows — exporter shares and concentration, not just a global correlation: the top exporter matches for 25 of 30 materials, exporter shares are within ~3.5% on average (median 3.2%), and the concentration ranking (HHI) correlates 0.92 (underlying flow structure: 0.975 log-value correlation across 21.7k flows). Absolute levels run ~1.8× higher (current Comtrade has absorbed revisions and late filers since BACI's early-2025 snapshot), so 2025 is calibrated to BACI's 2024 scale per material. A by-product finding: CIF/FOB rates are not identifiable on a 31-product slice (R²≈0.01) — BACI's gravity estimation needs the full ~5,000-product universe — so we fall back to a robust per-product median markup.

Caveat — partial coverage. Only about half of countries have filed 2025 annual data so far, so 2025 is genuinely provisional: stable monopolies (niobium, gallium) hold, but a material that leans on a country which has not yet reported (e.g. bauxite, whose dominant Guinea→China flow China has not filed) is distorted. Treat 2025 as indicative, not measured. Every earlier year (2018–2024) is reconciled CEPII BACI, untouched by this.

2026 (flagged 2026**) goes one step further — a directional scenario. The year is barely underway, so there is no annual data at all; only ~one quarter of monthly trade exists. Rather than invent a structure, it carries 2025's reconciled bilateral structure forward and scales each material by its Q1-2026 vs Q1-2025 export momentum (reporter-matched, from monthly Comtrade) blended with the change in its commodity price (World Bank Pink Sheet, for the ~7 base metals it covers). One quarter cannot credibly predict who-trades-with-whom, so shares are held at 2025 and only levels tilt: platinum rises (corroborated by a ~2× price move), antimony and graphite fall (export-control pressure). It is a trend signal — where each material is heading as of mid-2026 — not a measurement.

Validation & failure cases

The reconciliation engine (comtrade-reconcile, open + reproducible) is validated against official CEPII BACI on what this atlas shows — exporter shares and concentration, not a global level correlation.

The figures above are for exporters; importers validate similarly (2024: top-1 22/30, share MAE 4.2%, HHI corr 0.97) — both sides of the bilateral matrix hold. The validation is reproducible in CI from committed data (no API key) — see the engine repo.

Where it is weakest — surfaced, not hidden (2024):

• Top-1 exporter misses (5/30): fluorspar, graphite, arsenic, phosphate, tantalum — close second-place races on diffuse codes.
• Largest share errors: fluorspar 9%, graphite 7%, beryllium 5%.
• Concentration (HHI) misses: beryllium (0.53 vs 0.80), tungsten (0.28 vs 0.40), manganese (0.44 vs 0.33).

Full per-material 2024 result (red rows = top-1 miss):

Which findings are robust — a per-commodity status map

Three rounds of review added a lot of caveats (BACI is not truth; the engine understates some concentrations; thin codes swing wildly). Scattered, they bury the actual result. So we fuse them into one status per commodity (build_commodity_status.py, derived from the thinness screen, the engine-vs-BACI spread over 2002–2024, and the trend test). Draw headline findings from the “robust” rows only; read the rest as indicative. Of 31 tracked commodities: 21 robust, 4 spread-sensitive (a real but mild engine-vs-BACI gap), 2 shared-code (gallium/germanium can’t be separated in trade), 3 thin-fragile (graphite, beryllium, arsenic — too little trade to measure), and 1 engine-understated (cobalt, where the engine misses DRC’s real dominance).

What it is and isn't valid for. Valid: shares, ranks and top exporters. Concentration (HHI): reported as a spread, not a single number. The atlas engine is itself a full reconciliation — CIF/FOB deflation plus inverse-variance reliability weights on the two mirror reports (reconcile.py), the same family as CEPII BACI — not a plain average. An adversarial council pressed the right point: neither reconciliation is ground truth, so it is wrong to “correct toward BACI.” We now report concentration four ways per commodity (build_recon_envelope.py → out/recon_envelope.json): the two raw mirror reports — exporter-only (X/FOB) and importer-only (M/CIF) — and the two reconciliations (engine, BACI). A second review rightly rejected our first framing of this as the raw pair “bracketing the truth” with the reconciliations “inside”: two reports biased the same way (CIF vs FOB, confidentiality, re-exports) do not bound the truth, and because HHI is non-linear the aggregate need not fall between them — indeed BACI lands outside the raw pair for 24 of 30 commodities and the engine for 13 of 30 (2024). So the four views are a mirror-report spread — a measure of how uncertain a commodity’s concentration is — not a confidence interval, and no column is ground truth. The spread is wide: median ~0.07–0.09 HHI, up to 0.54 (manganese). The one directional thing we can say: the atlas engine tends to run low — often below both mirror reports (graphite 0.14 vs 0.23/0.48; beryllium 0.53 vs 0.62/0.78) — so its HHI is likely an underestimate for thin, dominated commodities. That is a direction, not a proven lower bound (an estimator can simply be wrong). Four ways, 2024 (exp / imp / engine / BACI): manganese 0.89 / 0.35 / 0.44 / 0.33, graphite 0.23 / 0.48 / 0.14 / 0.21, beryllium 0.62 / 0.78 / 0.53 / 0.80. The earlier one-number “beryllium 0.53→0.80” overstated false precision: read it as a wide, unresolved spread of roughly 0.5–0.8, the engine at the bottom and BACI at the top. Which commodities are too thin to carry a concentration number at all? Rather than judge that after the fact, we screen every code up front (build_thinness_screen.py): a commodity is flagged thin-fragile if its annual trade is under $100M or it has fewer than 8 active exporters. (High single-exporter leverage is not a flag on its own: in a deep market it is genuine one-country dominance — bauxite→Guinea, tungsten→China — a finding, not a data problem.) Only 3 of 31 are thin-fragile (graphite, beryllium, arsenic); their concentration is indicative, not measured. Building this ex-ante screen is also what caught our own cobalt error — see the status map above and the trend note below. Is the understatement fixable at source? We tried (reconcile/operator_test.py); simple operator substitutions do not remove it. The two-sided reconciliation’s geometric mean was the obvious suspect, but switching it to an arithmetic level-space mean changes the concentration almost nothing. The other diluter is importer-only flows (imports with no matching exporter report — often re-export/entrepôt) inflating the tail; but a full down-weight sweep shows no middle setting helps — partial weights barely move the BACI gap while worsening the fit, and only dropping them entirely improves raw-bracket coverage, at the cost of over-shooting past BACI. We also tested the principled candidate a reviewer named — the log-normal smearing correction (multiply the geometric mean by exp½σ² to retransform from log-median to level-mean). It fails badly here: with only two reports per flow the variance σ² is far too noisy, and the exponential over-inflates high-disagreement flows (a 100× mirror gap becomes a ~13× multiplier, pushing the value above both reports), so concentration over-shoots — raw-bracket coverage collapses from 17/30 to 6/30. We measure all of this two truth-agnostic ways: agreement with BACI (an external reconstruction, not ground truth) and coverage of the raw [exporter, importer] bracket; no operator we tested improves one without hurting the other. The one lever left untried is a gravity-fitted CIF like BACI’s, which reconcile.py documents as statistically unidentified on this 31-code slice (R² ≈ 0.01). So among the operators available to us, none improves the estimate — we report the spread as-is rather than hand-tune the engine toward BACI, which is not a target to hit. Does this band bias the trend claims? We had asserted “the dilution shifts the level not the trend” on only two years, which was not enough — so we did the full test (build_trend_robustness.py → out/trend_robustness.json): recompute the export-HHI series every year 2002–2024 on the engine reconciliation and independently on CEPII BACI, and run a Mann–Kendall trend test on each — and, because a reviewer noted annual HHI is serially correlated (which inflates plain-MK significance), using the Hamed–Rao autocorrelation-corrected variant. The correction barely moves the picture: it drops the engine’s significant trends from 21 to 20 of 29, so the series really are trending, not just persistent. A second honest point the reviewer forced: the raw “sign+significance agree for 26 of 29” headline is inflated because 8 of those agree only by both finding “no trend,” which is cheap. So we report the quantity that actually matters: among the 18 commodities where both series detect a significant trend, the direction agrees 18 of 18 — the reconciliation choice never flips a real trend’s sign, and all 18 survive a Benjamini–Hochberg multiple-testing correction. The three disagreements are all significance-level (one series clears the bar, the other just misses: arsenic, titanium), not direction reversals — except cobalt, the one substantive conflict (engine: a significant fall; BACI: a rise), which we ran down — and it exposes a real engine limitation, not a BACI one. (Correction, 31 Aug: our first pass called this a “thin-code artifact” on the strength of a buggy extraction that had dropped most exporters; cobalt’s code is not thin, and the story is the opposite.) Cobalt’s refined code (HS 2822.00, oxides/hydroxides) is a real $0.8–4.9 billion market, and BACI captures a genuine event: DR Congo’s share of those exports surged from 41% (2016) to ~85% (2020–2023) — the battery-era cobalt-hydroxide boom — then fell back in 2024 on the price crash (BACI HHI 0.25→0.71→0.17). The atlas engine missed it entirely: in 2020 it reconstructs only $0.7B of the trade (BACI: $3.5B) and puts DRC at 15%, not 84%. The reason is the engine’s own reliability weighting: DRC is a weak customs reporter, so its large but under-corroborated export claims get shrunk in reconciliation — exactly the “engine understates concentration” failure mode, here on a big, real dependency rather than a thin one. So the honest reading reverses: for cobalt, trust BACI and production (DRC ~70% of mine supply), not the engine’s smoothed series, whose “falling concentration” is an artefact of progressively shrinking DRC. This does not touch the 18/18 significant-trend agreement elsewhere, but it is a genuine scalp for the reconciliation engine, kept in view rather than buried. Not valid: absolute dollar levels — there is a known ~1.5–1.8× offset (current Comtrade runs above BACI's published values; diagnosed above), so arc and Sankey widths are relative, not dollar-precise. The earlier claim that the offset is a clean uniform multiplicative factor that “cancels in shares by construction” was too strong and is retracted: the two-year measurement above shows the offset falls unevenly on the leader, which is precisely why concentration needed correcting (reproducible: build_level_bias_audit.py, build_hhi_correction.py). 2025* is provisional (partial reporting, level-calibrated to BACI; not independently validated until BACI 2025 is released). 2026** is a directional scenario (Q1 momentum, shares held at 2025) — a trend signal, not a measurement.

Analytical layers & trend analysis

On top of the mine→refine→trade→reserves core, the atlas adds several screening lenses — each documented and caveated on its own page, and, with full citations, in the technical note:

• Trade-network chokepoints — directed-network betweenness / PageRank plus a node-removal fragility test (read as trade-routing centrality, since betweenness also reflects import size).
• Governance-weighted criticality — an EU/SCRREEN-shaped supply-risk proxy weighting concentration by World Bank governance scores, plus a Graedel-style axis; a proxy on public approximations, not the official scores.
• Concentration in tonnes (GeoPolRisk) — the atlas headlines concentration on trade value; this recomputes the Herfindahl on physical production tonnes (World Mining Data), weighted by WGI governance, aligned to the GeoPolRisk indicator (Gemechu et al. 2016; Cimprich/Helbig/Sonnemann 2017–2024) and EU CRM. It sets value, volume and production HHI side by side — trade value understates gallium (0.50 vs 0.97) and overstates boron (0.96 vs 0.28) — putting the concentration claim on the field’s standard physical footing.
• Economic complexity — Balassa revealed comparative advantage over the 32-material export matrix (specialization within critical materials), with ubiquity and relatedness.
• Origin trace — a first-order re-attribution of refiner-fronted imports to the dominant mine (an upper bound on single-origin concentration).
• Case studies — the atlas audited against five chains experts know cold (USGS / IEA).
• Scorecard uncertainty (Monte-Carlo) — propagates explicit input uncertainty through the five-axis hardest-cases scorecard (20,000 draws/material) to report P(hardest) as a gradient rather than a binary cutoff; confirms gallium/germanium are robust, surfaces vanadium as borderline, and shows supply elasticity + market thinness (not concentration) drive the ranking. The five input spreads are stated as judgement calls.
• Production in tonnes & cross-source check — absolute mine output (World Mining Data 2026, metric tonnes) for 28 materials, used to (a) show the physical scale behind the shares and (b) independently cross-validate the USGS-derived producer geography: a second authority names the same top producer for 26 of 28 materials (mean share gap 6.4pp), with the two genuine disagreements flagged, not smoothed.
• Demand & the squeeze — the demand-side arm: end-use sectors, clean-energy exposure and a demand-growth multiple to ~2040 (IEA Global Critical Minerals Outlook 2024 + USGS), crossed with companionality to separate demand-driven-but-mineable materials from the structural squeeze. Forward demand is scenario-dependent — round mid-scenario tiers, not forecasts. A companion price test checks the thesis against the atlas's own implied trade-price series (2018–24): by-product-locked metals split in direction by their host's cycle, not a mechanical price rise — the correlation is de-duplicated for the gallium/germanium shared HS6 code. A companion regression on real USGS price series then asks whether by-product metals are also more volatile: they are, but the gap is fully accounted for by market size, not companionality — small primary metals swing just as hard. A further host-coupling test correlates each companion's price against its host on real USGS prices, with a partial correlation removing the common commodity cycle and hosts pre-specified rather than chosen on the outcome. It refutes the "host decides" mechanism it was built to confirm: mean coupling 0.03, only bismuth←lead surviving, and nothing at all in the channel the theory actually names (the host's output, not its price). The episodes — gallium/germanium export controls, cobalt and Indonesian nickel — are real events and stand; the general law does not. Finally a demand-by-bloc read uses import flows as a revealed-demand proxy to show which bloc pulls each squeezed metal and the policy behind it — with the explicit caveat that imports conflate processing with final consumption (China's share is inflated by refine-and-re-export). A follow-up net-demand page nets that out (imports − exports), collapsing re-export hubs and revealing refiner-exporters (China is a net supplier of eight refined criticals) — a half-step toward apparent consumption, short only of absolute production tonnages. A supply-shock cascade then ties the whole arc together: a country shock is propagated through production shares, the companion web (a lost host drags its by-products down), and net trade to the exposed blocs — a structural first-and-second-order contagion map (no price feedback; echo only through tracked hosts), not a calibrated forecast.
• Gallium end-to-end (deep dive) — one metal followed physically from bauxite through alumina refining to recovery and export, with explicit bias bounds: a mass balance showing ~94% of the gallium passing through alumina refineries is discarded (recovery ~5.6%), so supply is capped by refinery retrofits, not price — the "can't scale" claim quantified rather than asserted, with the HS-811292 trade aliasing and China's non-reporting flagged.
• Supply structure — companionality & its children — a supply-elasticity arm: whether each material is mined for itself or recovered only as a by-product of a host commodity (companionality, from USGS MCS 2024 + Nassar et al. 2015). It spawns a supply-elasticity-adjusted risk, a host-shock map (the bulk commodities that gate critical supply), and a secondary-supply lens (EU CRM 2023 end-of-life recycling). Companionality is a round, literature-based estimate of a fuzzy quantity — read the axis as tiers, not decimals.

The reconciled trade series runs 2002–2024, and the Trends view does not merely plot it: each material's export-concentration, China-share and origin-gap series is tested with the Mann–Kendall trend test, the Theil–Sen slope and the Pettitt structural-break test, Benjamini–Hochberg-corrected across the 32 materials. 9 of 32 materials show a statistically significant rising export-concentration trend, with structural breaks clustering in 2012–2016. The critical-minerals literature typically plots concentration; testing and dating it is the contribution. Full methods and references: technical note.

Data & reproducibility

Global bilateral trade (map / globe / flow): primary source UN Comtrade, used in its reconciled, mirror-harmonised form CEPII BACI (Gaulier, G. & Zignago, S., 2010, BACI: International Trade Database at the Product-Level, CEPII WP 2010-23) — releases HS02 + HS17 V202601, years 2002–2024 (HS02 vintage through 2016, HS17 from 2017). The year is selectable in the tool. Does the 2016/17 vintage join manufacture a trend? Tested: the mean year-over-year change in top-exporter share at the splice (4.6 pp) is if anything smaller than a typical year (5.1 pp average across all boundaries) — −0.5 standard deviations, well inside ordinary variation. The splice is invisible in the series, so trends across 2002–2024 are not artefacts of it (reproducible: build_splice_sensitivity.py).

EU import-origin lens (Table view): Eurostat Comext, dataset DS-045409, extra-EU imports, annual. Base-R pipeline, parameterised by CN code, refreshes from the live API. Mine production & reserves: USGS Mineral Commodity Summaries (annual editions 2020–2024, with a per-layer year slider). Refining shares: best-per-material — BGS World Mineral Statistics (OGC API, annual, for the metals it reports) else IEA Critical Minerals Outlook 2026 / EU CRM 2023 / USGS. All sources are public and free (no API key); figures are rounded.