From 2 L to 2,000 L: what actually breaks when you scale a CHO process

Research 38 primary sources read, 36 cited: papers, vendor-authored studies and ICH Q5E 27 Sep 2026

3 to 7x

Ask why a CHO process misbehaves at 2,000 L and the first answer is often shear from the impeller. We read 38 primary sources, published between 2004 and 2026, to see what the evidence actually shows. The short version: turbulence from the impeller is rarely what kills cells. The documented trouble between 200 and 2,000 L clusters on the gas side (CO2 the vessel cannot strip, spargers that do not behave like the bench ones) and on the harvest step.

The five findings in short

• Impeller turbulence is not what kills cells at 2,000 L. A common explanation is that the turbulent eddies made by the impeller shrink to the size of a cell and tear it apart. By our calculation, at normal agitation the average eddy stays 3 to 7 times larger than a CHO cell. Lethal damage is reported where bubbles burst and at fast sparger jets, not in the bulk. Confidence: high.
• Impeller shear is sublethal, but it is real and clone-specific. One manufacturer measured titre losses of 10% or more at average shear stress levels that production vessels reach. Confidence: medium.
• Constant P/V is not a scale-up strategy on its own. You need a second criterion on the gas side, tied to your sparger. Confidence: high.
• pCO2 reaches product quality mostly through side effects. Its direct effect on glycans is small; the likelier route runs through base addition, osmolality and lactate. Confidence: medium.
• Harvest shear has cost real batches. It belongs in the scale-up plan, not after it. Confidence: medium.

How we built this

We read 24 sources in full text, 13 from the abstract only, plus the ICH Q5E guideline. Three sources were written by suppliers (Sartorius and Merck Life Science) and are marked as such. Two more have co-authors from equipment makers: SATAKE MultiMix on ^S01 and T&J Bio-Engineering on ^S05. Every figure in this piece is traced to a literal passage in its source. Where we computed something ourselves (eddy sizes, vessel geometry), the text and the figures say “BIOT calculation”. Each conclusion carries a confidence level: high means several independent sources agree; medium means the chain is built from different studies or one company’s data; low means two studies at most.

What changes when the volume goes up 1,000 times

Take a vessel family that keeps its height at twice its diameter, as the Sartorius single-use range does from 50 to 2,000 L ^S22. Going from 2 L to 2,000 L multiplies the diameter by 10. By our calculation the liquid column grows from about 0.22 m to 2.2 m, the liquid surface per litre drops 10 times, and the extra pressure at the bottom rises from about 2% to about 21% of atmospheric. Each of those has a documented consequence: less gas exchange through the surface, so more sparging is needed ^S02; weaker CO2 removal by surface aeration ^S30; and more dissolved CO2 at depth ^S03.

Agitation is the second change. Scale-dependent variables are interdependent, so they cannot all be held constant ^S02. By our calculation for a 10 times wider vessel, holding power per volume constant roughly doubles tip speed, and holding tip speed constant cuts power per volume to a tenth. Equal P/V can also hide very different shear: in 96-well plates, a CFD study cited by an RWTH Aachen team found shear rates up to 1,000 times apart at the same P/V ^S04. The Sartorius family, scaled at a constant kLa of 7.9 1/h, needed 162 rpm at 50 L and 70 rpm at 2,000 L ^S16.

Shear: why “eddies tear cells apart” misleads at scale

The classic worry is that turbulent eddies shrink to the size of a cell and tear it. The numbers do not support it for the bulk of the vessel. CHO cells are about 15 µm across ^S08. P/V is typically kept between 10 and 80 W/m3 ^S02, and a vendor design space from lab to production scale runs from 10 to 250 W/m3 ^S22. By our calculation, using the medium viscosity reported in a published CFD study ^S09, that gives average eddies of about 48 to 108 µm. Bringing the average eddy down to cell size would need roughly 30,000 W/m3.

A meta-analysis by Chalmers, cited by the Astellas team, places lethal responses such as apoptosis and necrosis mainly at 1,000,000 to 100,000,000 W/m3 ^S01. Those intensities do exist in a bioreactor, but locally: earlier simulations, reviewed by a Boston University and Biogen team, put bursting bubbles at 10,000 to 10,000,000,000 W/m3, and past studies found that most damage in sparged reactors comes from bubbles bursting at the free surface ^S21. The third source of damage is gas entrance velocity at the sparger. Boehringer Ingelheim saw up to 170 m/s at 12,000 L, against less than 1 m/s at 2 L, and proposes more than 60 m/s as critical for CHO; hole plugging made it worse ^S27. In perfusion, lysis came mostly from sparging, then from the hollow fibres, and least from agitation ^S34.

Astellas reached the same conclusion from the other side: across 10 parameters, average shear stress correlated best with titre (ρ = -0.64), while average Kolmogorov scales were “consistently much larger than the measured cell size” and correlated poorly ^S01.

That does not make the impeller harmless. The same Astellas work found that average shear stress of about 3.0 to 3.4 Pa was the threshold for a titre drop of 10% or more in a 3 L scale-down vessel. Its CFD put manufacturing stainless steel vessels at up to 6 Pa (3.02 Pa at 1.8 kL, 5.85 Pa at 5 kL), and it reports 1 to 2 kL single-use vessels also at up to 6 Pa, without showing the data. At 5.2 Pa, at least half of the antibody clones and 42% of the Fab clones lost 10% or more of titre ^S01. Earlier work by Sieck and colleagues, cited in a 2022 CFD study, reports productivity and stress-gene effects in CHO at an average energy dissipation as low as 0.4 W/kg ^S09. Sublethal, then, and clone-dependent.

Our reading: “shear” at 2,000 L is two separate problems. Lethal damage is a bubble and sparger problem. Titre loss from the impeller is a clone-selection problem, and a high-shear arm in the scale-down model can catch it before the clone is locked. Confidence: high for the first part, medium for the second, because the titre thresholds come from one company’s cell lines.

The gas side decides: CO2 and the sparger

Oxygen transfer grows more sensitive to agitation as the vessel grows: in UCB’s correlations the exponent on P/V in the kLa equation rose from 0.28 at 2 L to 0.55 at 10 L and 0.72 at 80 L ^S23.

CO2 behaves differently. In large vessels the gas stays in the liquid longer, so bubbles saturate with CO2 before they leave and stripping stops improving with kLa ^{S26 S02}. Takeda scaled a fed-batch process from 200 L to 2,000 L stainless steel at constant P/V and got higher dissolved CO2 at 2,000 L. Matching the ratio of CO2 to O2 transfer coefficients fixed it, and the corrected condition ran at a lower P/V than the 200 L vessel ^S28.

Constant P/V plus constant gas flow per volume (vvm) can work: a study cited by Lemire and colleagues reports comparable growth and titre at 3, 500 and 2,000 L that way ^S02. But vvm only transfers if the sparger does. Pore sizes in commercial single-use vessels range from 0.178 mm at 50 L to 0.582 mm at 2,000 L in one vendor’s line, 0.02 to 1 mm in another and 0.5 mm in a third, and using “the same vvm at different DHS pore sizes is still a flawed scale-up strategy” ^S05. Frit spargers give 4 to 20 times the kLa of drilled-hole spargers at equal vvm ^S13. Too much gas also hurts at small scale: in a 1 L study, a high air cap cut maximum titre from 1,617 to 1,228 mg/L through earlier cell death ^S02.

Our reading: scale-up needs two criteria. P/V or tip speed sets the liquid side. The second must be on the gas side and specific to the sparger: the kLa ratio of CO2 to O2, or a vvm chosen by pore size. Confidence: high.

pCO2: the quality damage comes through the side door

The thresholds are well mapped. In one CHO line the optimum for growth and productivity was about 90 mmHg ^S18. A rise from 50 to 150 mmHg cuts the specific growth rate by about 9%, inhibition can reach 30% at 250 mmHg, and a 2026 review recommends staying below 120 mmHg at large scale ^S14. At bench scale, dissolved CO2 above 68 mmHg has hurt growth and productivity ^S13. The opposite also happens: earlier work, cited by Lemire and colleagues, links sustained ultra-low pCO2 from heavy sparging to lower cell density and viability ^S02.

What is less obvious is where the product-quality damage comes from. A 2026 meta-analysis found that manipulating pH, dissolved oxygen or pCO2 rarely moved galactose, fucose or sialic acid content by more than 5%. Osmolality and temperature did: higher osmolality consistently lowered galactosylation by about 10% ^S11. And pCO2 feeds osmolality. CO2 evolution affects pH control, pCO2 and osmolality ^S10; in a Boehringer Ingelheim and University of Stuttgart study, a process without base addition ran at 357 mOsm/kg in stationary phase against 440 mOsm/kg with it ^S20. Elevated pCO2 also suppressed the lactate metabolic shift in TU Wien’s cultures ^S19, and early pCO2 shifts to 200 mbar gave a short productivity peak followed by 80% more lactate production per cell ^S20. At 2,000 L, Merck and Co. saw high CO2 and poor mixing restart lactate production, which drives more base addition; better CO2 removal broke the cycle ^S29.

Our reading: pCO2 is the failure mode to watch between 200 and 2,000 L, but measure what it does downstream of itself. Base consumption, osmolality and the lactate curve are the early signals, and they are what moves glycans. Confidence: medium, because no single study measures the whole chain at 2,000 L.

Mixing: mostly a problem above 5,000 L

A 3 L vessel mixes in about 10 seconds ^S06. The Sartorius single-use family reports under 30 seconds at every scale up to 2,000 L, a vendor figure ^S22. Large stainless steel is a different world: over 100 s at 5 m3, 120 to 360 s at 12 m3 ^S07, 100 to 170 s at 12,000 L depending on P/V in an earlier study cited by a Stuttgart team ^S24, and about 120 s at 15,000 L ^S06. Dissolved oxygen gradients were only about 10% even at 15,000 L ^S06, and bulk CO2 gradients are small next to those of oxygen or substrate ^S03. The sharp gradients are local: pH swings of up to 0.4 units were measured in an 8,000 L vessel with base added from the top ^S06.

When fluctuations do happen, they matter. In an AstraZeneca and UCL scale-down study, oscillating dissolved oxygen cut growth and titre by 35% in the fastest-growing line and raised lactate in all of them ^S07. In a study reviewed by a TU Munich team, a single multi-compartment scale-down vessel raised CHO lactate accumulation by 87% ^S03.

Our reading: at 2,000 L the mixing risk sits at the feed and base ports, not in the bulk. Subsurface or well-placed additions are the first lever to check. Bulk gradients become the main issue above 5,000 to 10,000 L. Confidence: medium, since the single-use mixing times are vendor data.

Harvest: where batches were actually lost

This is the finding with the clearest cost. During the scale-up of several CHO antibody processes, Genentech saw significant reduction of the interchain disulfide bonds, correlated with “excessive mechanical cell shear during the harvest operations”, and it led to failed specifications and lost batches ^S31. A 2026 study by Gilead, Bayer, Lonza and Ultragenyx describes the mechanism: shear- or pressure-induced lysis releases the cell’s thioredoxin and glutathione reduction pathways into the harvest. The authors built a bench assay that mimics the pressure of depth filtration at 200 to 2,000 L, typically run at 30 to 50 LMH ^S17. Merck Life Science recommends a pilot-scale step before going above 2,000 L ^S38.

Our reading: the scale-down toolkit should include the harvest, not stop at the bioreactor. Confidence: medium: two independent industrial groups, and the lost batches come from one company’s report.

Which failure dominates at which scale

Putting the evidence together gives a ranking that changes with volume. At 2 L the typical problems are overgassing, CO2 stripped too far, antifoam toxicity ^S04 and osmolality creeping up with bolus feeds (from 379 to 419 mOsm/kg in one study) ^S02. Between 200 and 2,000 L, CO2 accumulation and the lactate and base loop take over, along with sparger mismatch and harvest lysis. Above 10,000 L, gradients of pH, oxygen and substrate join the list ^S24; Nienow already flagged pH and nutrient homogeneity at the largest scale in 2006 ^S10. Impeller shear sits in the background at every scale, sublethal and clone-dependent. Confidence: medium.

Tools that transfer, and tools that do not yet

• Scale-down models matched to the failure mode. A baffled 3 L vessel that reproduces up to 5 kL shear ^S01, gas entrance velocity mimics ^S27, oscillation studies in ambr15 ^S07 and single multi-compartment vessels that reproduce large-scale mixing times ^S24. One caveat: models that fake pH swings with bolus base can push osmolality to 550 mOsm/kg, an artefact of the model itself ^S06.
• Poloxamer 188. Without it, CHO could not grow under direct sparging ^S33. It is typically used at 1 g/L, and Amgen reports industry experience with variability in the shear protection it gives ^S32.
• Soft sensors across scale. In a Sartorius study, capacitance tracked viable cell count during exponential growth with an R2 of 99.2% at 50 L and 99.0% at 2,000 L ^S16. Several Raman models built in Ambr 250 mini-bioreactors transferred poorly to 5 L bench vessels; the authors, citing earlier work, note that such failures often come from differences in instruments, probes and acquisition settings. Tailored preprocessing cut the cross-scale error by 14.0 to 56.1% ^S12. Our reading, with low confidence (two studies): physical signals transfer more easily than spectral models.
• Models. Transient CFD coupled to cell kinetics “is not possible currently” for mammalian fed-batch ^S08; compartment models are the practical route. A hybrid approach transferred a mechanistic model between systems with a single dataset, tested so far from shake flask to 1.5 L ^S15. The largest scale-up dataset we found, compiled from published data, has 18 processes from 55 bioreactors between 250 mL and 5,000 L ^S25.
• Regulatory frame. ICH Q5E lists increasing scale among the process changes that call for comparability, and for approved products commercial-scale data are “generally indicated” ^S36.

What it means for a plant

• Pick two scale-up criteria, not one: P/V or tip speed for the liquid, and a gas-side criterion (CO2 to O2 kLa ratio, or vvm by pore size) for the sparger you will actually use at scale.
• Get the sparger specification of the production vessel before the scale-down runs. Pore size and hole count decide gas entrance velocity and CO2 removal.
• Track base consumption and osmolality as closely as pCO2. They are the route by which CO2 reaches product quality.
• Add a high-shear arm to clone selection; at least half of the antibody clones in one study lost 10% or more of titre at 5.2 Pa.
• Scale down the harvest: depth filtration pressure and lysis belong in the same plan as the bioreactor.
• If you move a Raman model from a mini-bioreactor to a larger vessel, budget for re-preprocessing or recalibration. Capacitance held up from 50 to 2,000 L in the one vendor study we found.

What we could not verify

We found no primary data on heat removal as a limit at 2,000 L for CHO, on foam-out frequency at 2,000 L, on disc-stack centrifuge shear, or a measured pCO2 time course at 2,000 L in open text. The common claim that local energy dissipation near the impeller is 10 to 100 times the average also lacked a primary figure. We left all of these out. Suppliers’ application notes for their 2,000 L vessels could not be retrieved for this piece. If you hold data on any of these, the corrections page is open.