Soft lithography is the standard way to make a microfluidic chip, and it has been for two decades. It is also slow, expensive, and unforgiving. This page walks the whole process, from bare silicon wafer to a bonded device, and is honest about where it goes wrong.
The process splits in two. First you build a mold, which is the part that needs the cleanroom. Then you cast against it, which is the part that needs the hands. A mold can be reused, so the first half is amortized. The second half is repeated for every single chip.
A bare silicon wafer is solvent cleaned and dehydration baked so the photoresist will adhere.
Any particle left behind becomes a defect reproduced into every chip cast from this mold.SU-8 is poured onto the wafer centre and spun to depth. Spin speed sets channel height, so the recipe is the geometry.
Edge bead, bubbles, or a speed error changes channel height across the whole wafer.Baked in stages on a hotplate, typically around 65 °C then 95 °C, to drive off solvent before exposure.
Under baking leaves solvent and the resist sticks to the mask. Over baking cracks it.The photomask is aligned to the wafer in a mask aligner and the pattern is written with UV.
Alignment is done by an operator through optics. Dose errors under or over cure the features.Uncured resist is dissolved away, leaving raised channel features. The mold is then vapour silanized so cured rubber will release from it.
Skip or rush silanization and the first casting tears the mold apart, taking the design with it.Base and curing agent are mixed, conventionally around ten to one. The ratio sets stiffness and how well it will bond later.
Mixing traps air throughout the batch, which is why the next step exists.The mixture sits under vacuum until the bubbles stirred in during mixing rise out and collapse.
A bubble that survives into a channel wall is a scrapped chip, found only after curing.Poured over the mold and cured with heat, commonly around 60 °C for several hours or overnight.
This is the wait. The design is frozen for the length of the cure before anyone learns whether it worked.The cured slab is cut out and peeled off the mold by hand.
The most physical step. Peel wrong and you tear the slab, damage fine features, or crack the wafer underneath.Every inlet and outlet is punched individually with a biopsy punch, positioned by eye against the channel end.
A punch that misses, tears, or leaves a plug of debris in the channel ruins the chip after all prior work.The slab and a glass slide are plasma treated, then pressed together to form a permanent seal.
One attempt. Contact is irreversible, and the activation is already decaying (see below).Cost is the number people quote. Time is the one that actually shapes the research, because it sets how many times a question can be asked and answered in a year.
The top two rows of that chart describe labs that can afford a cleanroom. Most cannot. The equipment and the room together are a capital commitment few research groups can justify for a fabrication step that is not itself the research, so the work goes out to a foundry, and the group joins the bottom row.
That is the part worth sitting with. A researcher with an idea on Monday is not asking when the chip will be finished. They are asking whether the question is worth a purchase order and a month. Ideas that would take an afternoon to test never get tested, because the fabrication step prices them out of existence.
And it compounds with iteration. A first design is rarely right. The second version means a new photomask, a new master mold, or a new order, and the same wait again. A group that could run ten iterations a year on a fast loop runs two or three on a slow one, and the difference is not effort or funding. It is fabrication turnaround.
Eleven steps, and roughly half of them depend on an individual's hands and judgement rather than on a setting you can dial in. How fast to peel. When the degassing is finished. Where exactly to place the punch. How much pressure to apply as two surfaces meet, once, permanently.
The consequence is that outcomes track the operator. A lab with someone experienced produces working chips. The same protocol, same equipment, and a new graduate student produces a stack of scrap while they learn the feel of it. That knowledge lives in people, not in documentation, which is why the capability concentrates in a small number of institutions and why it does not transfer easily between them.
It also shows up in the literature. Even the standard test for whether a bond held is a manual peel, which yields no quantitative data and is acknowledged to be subjective with poor reproducibility. The field's own quality check is a person pulling on a chip to see what happens.
Most of these are only discovered at the end, after every prior step has already been paid for in time and material.
The classic failure. A bond fails when a path of separation can be traced from the channel out to the edge of the device. Under pressure, the interface is what gives.
Tubing pushed into a hand punched port pops back out under flow, because the port was punched slightly oversized or off axis.
The rubber itself tears rather than the bond, which is what happens when the mix ratio was adjusted to favour bond strength at the cost of material strength.
Air that survived degassing sits in a channel wall. Invisible until the slab is cured and cut, at which point the chip is scrap.
Peeling a cured slab off the master is done by hand. Fine or high aspect ratio features are the first to go, and a badly stuck slab can take the mold with it.
Alignment takes longer than expected, the activated surface decays, and the bond that results is weak in a way that only shows up later under flow.
Those first three are documented failure classes in the bonding literature, not anecdotes: visible delamination, inlet pin ejection, and bulk PDMS splitting are the named modes researchers sort their broken devices into.
Everything above is a fabrication defect. This one is not, and it survives a perfect build. PDMS is biocompatible in the ordinary sense, and medical grade silicones are used in implants, so the issue is analytical rather than toxicological. PDMS is a porous, hydrophobic polymer network, and small hydrophobic molecules partition out of solution and into the bulk of the material. Toepke and Beebe reported more than 90 percent of a small hydrophobic dye absorbed into the channel walls within 24 hours. For a drug dose response or a pharmacokinetic study, that means the concentration reaching the cells is lower than the concentration that was added, and it keeps falling while the experiment runs. Uncrosslinked oligomers travel the other way, leaching out of the bulk and into the medium. Neither effect presents as a broken chip, which is what makes it easy to miss, and it is a large part of why groups doing assay work move to thermoplastics such as COC.
Material class is the lever. Uptake of this kind is characteristic of elastomers, which are loosely crosslinked and have high free volume. Glassy materials, a group that includes thermoplastics such as COC and cured printed resins, show mainly surface adsorption instead of bulk absorption. Lower uptake is therefore an expected benefit of a resin based chip, and it is one we have not yet measured against a PDMS control.
Absorption as reported by Toepke and Beebe, Lab on a Chip, 2006, with a broader review of partitioning behavior in Micromachines, 2022.
There is a second, quieter version of the same problem. The plasma treatment that makes bonding possible also makes the surface temporarily water friendly, and that property does not last either. The same hydrophobic recovery that closes the bonding window keeps going after the device is sealed, so the internal surface a protocol was validated against is not necessarily the surface it has weeks later.
Neither issue is exotic or a sign of a bad lab. They are properties of building a channel by sealing two parts together, in a material chosen for how easily it casts.
Two very different answers depending on scale. A research bench is affordable to a funded lab. A production line is a different category of purchase entirely, and it is the reason the field has not simply scaled its way out of the problem.
| System | What it does | Price |
|---|---|---|
| Mask aligners SUSS MA6, MA8, MA150, MA200; EVG 620, 6200, Smartview |
Aligns the photomask to the wafer and exposes the pattern | $150k – $500k+ estimate |
| Wafer bonders SUSS SB6, SB8; EVG 501, 520, Gemini |
Bonds substrates under controlled thermal compression | $150k – $500k+ estimate |
| Automated coater track | Automated resist dispensing and coating | $150k – $400k estimate |
| PVA TePla 660 | Low pressure microwave plasma cleaning | $35,000 used listed |
| PVA TePla ION-100 | Plasma surface treatment, cleanroom ready | $29,500 used listed |
| THINKY ARE-400TWIN | Planetary centrifugal mixing and defoaming at volume | $20k – $35k estimate |
None of this equipment carries a published price, and that is the point. ClassOne Equipment lists a full catalogue of refurbished SUSS and EV Group mask aligners and wafer bonders, covering MA6, MA8, MA150, MA200, EVG 620, 6200, Smartview, SB6, SB8, EVG 501, 520, and Gemini. Every one is quote only, with no figure attached. Capital equipment in this class is sold on inquiry, which is exactly why the ranges above are marked as estimates: they are market ranges rather than published prices. The two marked as listed are the exception, verified against current listings at BidService. Accessories are capital items too: a replacement wafer chuck or maskholder runs roughly $3,000 to $6,500.
| Item | What it does | Price |
|---|---|---|
| Mask aligner | Aligns and exposes through the photomask | ~$50,000 |
| Mask exposer | UV exposure without full alignment optics | ~$15,000 |
| Plasma machine | Surface activation before bonding | ~$7,000 |
| Spin coater | Sets photoresist thickness, and so channel height | ~$7,000 |
| Consumables | SU-8 photoresist, PDMS, chrome photomasks | $400/L, $200/kg, $100–500 |
| Typical full bench | Conventional soft lithography setup, equipment only | $30,000 – $65,000 |
Research bench figures are as published in Low-Cost, Accessible Fabrication Methods for Microfluidics Research in Low-Resource Settings (Micromachines, 2018), a peer reviewed survey of conventional versus low cost approaches.
Soft lithography is how chips get prototyped. Injection molding is how they get produced. It solves the throughput problem and creates a different one, because the geometry stops being a file and becomes a piece of machined steel.
| Item | What it does | Price |
|---|---|---|
| Aluminium tool | Prototype and low volume runs, fewer cycles before wear | $5,000 – $30,000 |
| Steel tool, simple part | Production tooling for a straightforward geometry | $15,000 – $50,000 |
| Steel tool, multi cavity | High volume production of a complex device | $50,000 – $100,000+ |
| Machining labour | CNC, EDM, grinding, polishing; 40 to 50 percent of tool cost | 80 – 200+ hrs |
Tooling figures as published by Meridian Medical and Formlabs. The number that matters is not the first tool, it is that you buy another one for every design change. A revision is not a file edit, it is a new block of steel and the weeks of machining that go with it. That is why labs prototype in PDMS and manufacture in thermoplastic, and why the device that gets validated is not the device that ships. Medical grade polymers and cleanroom molding are priced on top.
Only the first half of the process needs a cleanroom. Photolithography on the silicon wafer is where a single airborne particle becomes a permanent defect in the master, and every chip ever cast from that master inherits it. Casting, curing, and bonding do not need the same air. So the room is a fixed cost carried by the mold step alone, and it still gates everything downstream, because there are no chips without a master.
How clean the room has to be is set by the finest feature on the master, not by the material. Fine geometry is typically run at ISO 5 to ISO 6, the classes also used for semiconductor and implantable device work. Channels in the tens of microns, which covers much of ordinary microfluidic work, are commonly run at ISO 7, and less demanding geometry is made in a standard laboratory with a laminar flow hood over the exposure steps. Across that range the room costs $250 to $1,000 per square foot to build before any instrument is installed. HVAC and filtration are 35 to 55 percent of that budget, which is also where the recurring cost sits, in energy, filter replacement, maintenance, and recertification.
Class limits per ISO 14644-1. Build cost as published by Labs USA (ISO 7 to ISO 5, $250 to $1,000 per sq ft, prefabricated) and Terrapin Consulting Group (ISO 7 at $400 to $650 per sq ft, constructed, 2026). Both note that final cost depends on build method and fit out. What a cleanroom is, what it does, and who needs one →
Add it up and the pattern is the same at every scale. The equipment is expensive, the room the mold is made in is expensive, and both are fixed costs that have to be carried whether or not anyone is producing chips that week. Underneath all of it sits the operator, who has to be trained, retained, and paid, and without whom none of the above produces a working device.
Descriptions only go so far. These are walkthroughs of the fabrication process filmed in working labs. They open on YouTube in a new tab.
Soft lithography is a good process that has carried the field a long way. It is simply a process built around a mold, a cleanroom, and a pair of trained hands, and every one of those is a constraint on how fast a lab can go from an idea to a part. We took out the step that forced the tradeoff.
See what we built instead