Where pre-FEED assumptions break between pilot and fab
A bench- or pilot-scale UPW train that shows 18.2 MΩ·cm resistivity and below 1 ppb DOC on a single-pass test is not the same machine as a 10–20 MGD fab UPW train, and the difference shows up first in TOC, recovery rate, reagent residuals, and material inertness (per sciencedirect.com 2025-08; hydropurewater.com 2026). A published pilot combining ultrafiltration with two-stage reverse osmosis can reach above 75% recovery and deliver RO permeate that subsequently produces 18.2 MΩ·cm resistivity with below 1 ppb DOC, yet small fractions of low-molecular-weight neutral substances such as urea and metal-humic complexes with chromium and nickel persist past both RO stages (per sciencedirect.com 2025-08).
The fab design envelope is tighter: 18.2 MΩ·cm at 25°C, TOC below 1 ppb at current nodes (below 0.5 ppb at 2 nm and 3 nm nodes), particles above 0.05 µm below 0.3 per mL, dissolved oxygen below 1 ppb, and bacteria below 1 CFU per 100 mL (per hydropurewater.com 2026 and SEMI F63 2026 revision). Net withdrawal of 2–10 MGD after 30–80% recycling means the recycle assumption is the single most fragile pre-FEED input (per hydropurewater.com 2026, citing SemiconductorX 2026).
The pilot cannot reproduce the residence time, biofilm surface area, and distribution-loop recontamination seen in a fab-scale UPW distribution loop, and pilot success on resistivity and DOC at low flow is necessary but not sufficient for fab qualification (per hydropurewater.com 2026). UMC's published process-water recycling rate of 84.3% companywide and ASE's Kaohsiung dedicated recycling plant at up to 30,000 tons per day at about 75% recovery are the realistic upper-bound benchmarks for what segregation and reuse infrastructure can deliver at production scale (per semiengineering.com 2026). The 2026 working assumption of 65–75% recycle with a documented glide path to 90% is the correct pre-FEED range, not the 84.3% headline (per hydropurewater.com 2026, citing SemiconductorX 2026). UMC's "new fab areas" route wastewater into as many as 27 categories to avoid co-mingling — pilots rarely run this many segregated streams, so the recycle fraction the pilot claims is not the recycle fraction the fab will achieve without engineered segregation from day one (per semiengineering.com 2026). For an engineer who has sold a 75% pilot recovery number internally, that gap is the first place the assumption breaks.
Pilot numbers vs. fab numbers: the parameter gap
The table below aligns pilot evidence with the 2026 fab design envelope so a process engineer can see which numbers travel from bench to fab and which ones do not.
| Parameter | Pilot evidence (UF + 2-stage RO) | Fab design envelope (leading-edge 300 mm, N3–N5) |
|---|---|---|
| Recovery | Above 75% with intermittent 2nd-stage RO to mitigate fouling | 30–80% gross recycle, per-pass RO 70–85%; 65–75% working assumption, glide path to 90% |
| Resistivity | Lab UPW from total RO permeate: ≥18.2 MΩ·cm | 18.2 MΩ·cm at 25°C (theoretical maximum for pure H₂O) |
| DOC / TOC | 0.5 mgC/L from RO permeate; lab UPW below 1 ppb DOC | Below 1 ppb at current nodes; below 0.5 ppb at 2 nm/3 nm nodes |
| Contaminant removal list | Low-MW DOM, urea, TMAH, PFAS, SiO₂; persistent species: protein-like low-MW neutrals, metal-humic complexes with Cr²⁺ and Ni²⁺ | Particles above 0.05 µm below 0.3 per mL; dissolved O₂ below 1 ppb; bacteria below 1 CFU per 100 mL; sub-ppb metals |
| Reagent residual | Not routinely tracked at pilot | 185 nm UV oxidation generates H₂O₂ detrimental to production; peroxide carryover into the polishing loop must be controlled |
| Operating mode | Intermittent 2nd-stage RO to enhance recovery | Intermittent 2nd-pass RO must be locked in pre-FEED — it sets CAPEX, footprint, and brine routing |
| Drain segregation | One or two combined streams in pilot | Up to 27 segregated drain categories (rinse, acid, alkaline, CMP, RO concentrate) |
Two cells in that table drive most of the FEED-scope delta. The first is the TOC row: pilot lab UPW from RO permeate clears below 1 ppb DOC, but a 2 nm or 3 nm fab requires below 0.5 ppb TOC, and the polishing loop, not the RO train, is what has to absorb that tightening (per hydropurewater.com 2026, citing AXEON 2026). The second is the intermittent 2nd-stage RO cell: the pilot uses it to enhance recovery and mitigate membrane fouling, and at fab scale this is the operating mode that has to be locked in pre-FEED because it changes CAPEX, footprint, and brine routing (per sciencedirect.com 2025-08). The fab contamination budget also adds three lines the pilot does not measure against: particles above 0.05 µm below 0.3 per mL, dissolved O₂ below 1 ppb, and bacteria below 1 CFU per 100 mL (per hydropurewater.com 2026, citing SEMI F63 2026 revision).
Assumption failures that show up first at fab scale-up
The ordering below is diagnostic: it lists the assumptions that fail first, the symptom each one produces, and the evidence in the supplied research that flags it. Engineers who have only the pilot data should read this as the failure-mode list their FEED scope has to cover.
- Organic trace accumulation. Low-MW neutral substances and metal-humic complexes that pass two-stage RO in the pilot (per sciencedirect.com 2025-08) are the precursor to fab-scale TOC excursions and earlier retirement of polishing-loop water; semiengineering.com 2026 notes that "as reuse fractions rise, trace neutrals and very small fragments can accumulate in ways that are hard to see with routine metrology."
- Reagent residuals. 185 nm UV oxidation produces hydrogen peroxide that is detrimental to production (per semiengineering.com 2026); pilot TOC reduction looks clean on a bench meter but a fab must account for peroxide carryover into the polishing loop, which forces a CO₂ degassing stage and a residual peroxide specification that the pilot never had to set.
- Material extractables. Tanks, liners, valves, gaskets, and distribution piping cannot add extractables back into the loop or recycle potential collapses; if a sump coating or tank lining leaches low-MW organics, the site retires partially spent water earlier than planned (per semiengineering.com 2026). Low-extractable, low-permeability linings are required to keep reclamation viable, and that is a yield decision, not just a durability decision.
- PFAS and persistent species detection. Trace neutrals and persistent species accumulate as reuse climbs, and PFAS detection limits are at or beyond parts-per-trillion, so pilot instruments miss what the fab will see (per semiengineering.com 2026). The practical takeaway is to treat purity as a plant-wide constraint and to spec pre-treatment or destruction at the point where the risk first appears.
- Drain segregation discipline. UMC's "new fab areas" route wastewater into as many as 27 categories to avoid co-mingling (per semiengineering.com 2026); pilots rarely run this many segregated streams, so the recycle fraction the pilot claims is not the recycle fraction the fab will achieve without engineered segregation from day one.
- TOC tightening at 2 nm / 3 nm nodes. Organic compounds contaminate EUV optics and interfere with gate dielectric formation, so the spec moves from below 1 ppb to below 0.5 ppb (per hydropurewater.com 2026, citing AXEON 2026). A pilot sized for current-node TOC will not qualify a 2 nm fab; the polishing loop has to be sized for the node on the FEED horizon, not the node in production today.
For engineers building the polishing loop side of the train, the calibration of resistivity, TOC, and DO trip points against a contamination budget is the topic covered in the guide on polishing loop distribution technologies for 18.2 MΩ·cm UPW.
Site, watershed, and recycle assumptions that change the train
Pre-FEED recycle targets are site-dependent, not industry-universal, and the watershed risk class changes the recovery design. TSMC Arizona Fab 21 sits on the Colorado River and Salt/Verde system with Lake Mead and Lake Powell at historic lows, and TSMC has committed to 100% recycling with ZLD on cooling-tower blowdown (per hydropurewater.com 2026, citing SemiconductorX 2026). Intel Ocotillo at full build draws about 14 MGD across three fabs — Fab 52 = 4 MGD, Fab 62 = 5 MGD, Fab 42 = 3 MGD — with about 4 MGD potable and 10 MGD reclaimed, reducing the potable burden by more than two-thirds (per semiengineering.com 2026, citing Intel EA). Phoenix's designated-provider framework requires demonstrated long-term supply out to a 100-year build-out, and Phoenix expects an industrial reclaimed water plant with RO upstream of discharge before anything reaches the city sewer (per semiengineering.com 2026, citing City of Phoenix Water Services).
Intel Oregon and Micron Clay sit in low-stress watersheds and can target above 80% reuse; Samsung Taylor in central Texas is moderate (per hydropurewater.com 2026, citing SemiconductorX 2026) — the recycle number has to be watershed-class-specific, not a corporate average. Engineers who lock a single recycle number into the pre-FEED package without naming the watershed class are setting up the project to fail a regulatory or community-relations review that arrives before the FEED budget is approved.
Cost, lead time, and FEED-window pressure
Assumption validation has to land in pre-FEED because late changes blow the procurement clock and the budget. UPW system CAPEX sits in the $200–500M range for a leading-edge fab, with the recycle system adding $50–150M on top, and RO membranes, EDI stacks, and UF modules carry 12–18 month lead times on US projects in 2026, so the procurement clock starts during site due diligence rather than after groundbreaking (per hydropurewater.com 2026, citing SemiconductorX 2026). The table below maps the assumptions that drive procurement clock and budget to the engineering package that has to absorb them.
| Pre-FEED assumption | Procurement / budget consequence |
|---|---|
| Recycle target as a point number | Forces a redesign when cooling-tower cycles of concentration and tool heat load move the operating envelope; design for 65–75% working with a glide path to 90% |
| Municipal feed ratio assumed flat | Municipal feed runs 1,400–1,600 gallons of city water per 1,000 gallons of UPW produced (per hydropurewater.com 2026, citing AXEON 2026); missing seasonal variability under-sizes the pretreatment train |
| RO membranes and EDI stacks as catalog items | 12–18 month lead times on US projects in 2026 mean the procurement clock starts in site due diligence; late spec changes slip groundbreaking |
| CAPEX modeled at primary train only | Recycle system adds $50–150M on top of the $200–500M primary UPW system; both are long-lead items in fab construction scheduling |
The municipal feed ratio is the silent budget driver: 1,400–1,600 gallons of city water per 1,000 gallons of UPW produced is the published envelope, with the difference lost to RO reject, cooling tower drift, and backwash, so a pre-FEED feed-water profile that misses seasonal variability will under-size the pretreatment train (per hydropurewater.com 2026, citing AXEON 2026). On the equipment side, the train is anchored by an industrial RO system sized for two-pass operation and by EDI polishing stacks for the 15–17 MΩ·cm output stage.
Pre-FEED checklist before the FEED budget is approved
The list below is the practical output of the article: six items a process or facilities engineer can put in front of a project sponsor before FEED is signed. Each item is grounded in the supplied research.
| Item | Why it has to land pre-FEED |
|---|---|
| Lock the source-water profile and seasonal variability before sizing the train | Municipal supply and reclaimed supply present different feed-water profiles; missing seasonality under-sizes pretreatment (per hydropurewater.com 2026, citing SemiconductorX 2026) |
| Recycle target as a range with a glide path (65–75% working, path to 90%) | Point numbers fail when cooling-tower cycles of concentration and tool heat load move the operating envelope (per hydropurewater.com 2026, citing SemiconductorX 2026) |
| Segregate drain streams (rinse, acid, alkaline, CMP, RO concentrate) into the fab drain from day one | UMC's 27-category routing is the operating evidence that segregation drives fab-scale recycle (per semiengineering.com 2026) |
| Specify ZLD on cooling-tower blowdown for any water-stressed site; reserve space for a third RO train on concentrate | Standard at Arizona sites; allows the recycle rate to climb without a shutdown (per hydropurewater.com 2026, citing SemiconductorX 2026) |
| Validate low-MW neutral and metal-humic complex removal past two-stage RO using site-specific wastewater | Urea and Cr/Ni complexes persisted in the published pilot; analogue feed will not reproduce the failure mode (per sciencedirect.com 2025-08) |
| Plan PLC-based continuous monitoring for resistivity, TOC, particles, and DO with alarm setpoints tied to the contamination budget | Pilot instrumentation is not fab instrumentation; metrology limits are at or beyond parts-per-trillion for PFAS (per semiengineering.com 2026) |
On the pretreatment side, a multi-media pretreatment filter sized for the seasonal envelope ahead of two-pass RO is the conventional US install, and consumables such as RO and UF membrane replacement elements are best specified as a separate procurement package so lead time is visible against the 12–18 month component window. For engineers connecting this to a broader plant water balance, the engineering guide on semiconductor process wastewater engineering and the reference on industrial water consumption reduction cover the upstream and downstream framing.
The recycle assumption deserves a direct callout. Pilot reports quote a single number because pilots run a single feed at a single flow; a fab runs five segregated drain streams through a coupled water-and-energy loop whose operating envelope moves with cooling-tower cycles of concentration and tool heat load. Treat the recycle target as a range with a glide path, not as a single number the pilot sold internally.
Frequently Asked Questions
What is the right pre-FEED recycle range to budget against for a leading-edge fab?
Plan for 65–75% as the working assumption and design the loop to climb toward 90% without major retrofit, per the 2026 fab design envelope in hydropurewater.com 2026 citing SemiconductorX 2026. UMC's published 84.3% companywide rate and ASE Kaohsiung's about 75% recovery on a 30,000 tons-per-day plant are the upper-bound benchmarks, not the floor (per semiengineering.com 2026). The recycle number has to be watershed-class-specific: 100% with ZLD at TSMC Arizona Fab 21, above 80% at Intel Oregon, moderate at Samsung Taylor (per hydropurewater.com 2026, citing SemiconductorX 2026).
What CAPEX line item and lead time should a project sponsor expect for the UPW and recycle system?
UPW system CAPEX sits in the $200–500M range for a leading-edge fab, with the recycle system adding $50–150M on top, and RO membranes, EDI stacks, and UF modules carry 12–18 month lead times on US projects in 2026 (per hydropurewater.com 2026, citing SemiconductorX 2026). Request from each bidder a procurement-package list that separates membranes, EDI stacks, and UF modules as long-lead line items with named manufacturers, so lead time is visible against the groundbreaking date rather than buried in the integrator scope.
Which pilot assumptions most often fail to translate to a 10–20 MGD fab train?
Six assumptions fail first: organic trace accumulation past two-stage RO (urea, metal-humic complexes with Cr and Ni per sciencedirect.com 2025-08); reagent residuals from 185 nm UV oxidation (H₂O₂ detrimental to production per semiengineering.com 2026); material extractables from sump coatings and tank linings (per semiengineering.com 2026); PFAS detection limits at or beyond parts-per-trillion (per semiengineering.com 2026); drain segregation discipline (UMC's 27-category routing per semiengineering.com 2026); and TOC tightening from below 1 ppb to below 0.5 ppb at 2 nm/3 nm nodes (per hydropurewater.com 2026, citing AXEON 2026). Validate each one with site-specific wastewater before FEED closes.
How do I scope a supplier shortlist for a fab-scale UPW train so the bid is comparable?
Ask each supplier for three inputs in writing: a watershed-class recycle range with a documented glide path, not a single recycle number; a procurement-package list naming long-lead items (RO membranes, EDI stacks, UF modules) with manufacturers and 12–18 month lead-time commitments; and a metrology package that ties resistivity, TOC, particle, and DO alarm setpoints to the contamination budget, with detection limits stated for PFAS at parts-per-trillion. Suppliers who answer with a single CAPEX figure and a generic lead time are not yet scoped to fab scale.
