A graphene membrane removes half the caffeine from coffee

A coffee filter that counts molecules

Decaf coffee sounds like a shelf product. In the plant it is usually solvent chemistry or pressurized water and a long cycle. Nanofiltration promises something else: less energy, a shorter path, and molecule-level selection in coffee liquor.

A team from UNSW Sydney and Okayama University (including Yuta Nishina) checks whether graphene oxide membranes can be tuned to cut caffeine while keeping compounds tied to coffee flavour and matrix. The paper in Journal of Membrane Science (DOI: 10.1016/j.memsci.2026.125992) does not announce “caffeine-free coffee.” It shows a controlled drop: from 24.64 to 12.95 μg/mg, about 47.4 percent less caffeine in a filtration test. At the same time, after filtration — depending on which side of the membrane — measurable HMF, trigonelline, NMP and choline remain.

That is a classic TRL09 reverse. A lab win is not yet a product. Half the caffeine gone is not zero. Retained flavour markers are not a guarantee that the brew “tastes as before” in an industrial tasting. But the direction is concrete: filtration instead of solvent, a membrane ranking instead of a single “how much went through” curve.

Why decaffeination is not simple

Caffeine does not sit alone in coffee. Around it is a dense matrix: chlorogenic acids, Maillard products, alkaloids, nitrogen compounds. Classic solvent decaffeination works, but it carries energy cost, chemical logistics and residue questions. Water and supercritical CO₂ also have their process price.

Nanofiltration (NF) is described in the literature as a lower-energy route to instant coffee with reduced caffeine. The condition is strict. The membrane must at once:

remove caffeine,
pass water with sensible permeance,
retain or protect key matrix compounds so the liquor does not become “empty.”

These three goals pull in different directions. Wider channels give flow but lose selectivity. Narrower channels hold more but throttle throughput. Adsorption on channel walls and confinement add further effects — a small surface-chemistry change can flip the result.

For industry, coffee is not only a caffeine percentage on the label. It is also trigonelline, roast products, sensory profile and batch repeatability. Hence this paper has no single “golden” number. It has a set of indicators.

Reverse one: graphene oxide dislikes simple rules

Graphene oxide membranes (GOMs) have a layered nanochannel structure. They can be modified by intercalating sodium alginate between layers, crosslinking with calcium ions, or combining both strategies. Intuition suggests monotonicity: “more modification = better.” The results show the opposite.

Modification effects are non-monotonic. Small structural and chemical changes create competing effects on permeance and selectivity. Transport, adsorption and confinement act together. Sometimes improving one KPI hurts another. That is the materials reverse: you cannot “tighten” the membrane with one knob and get everything at once.

The authors do not pretend they found a universal recipe. They build an evaluation frame meant to choose a compromise — not an absolute winner in a single category.

Five KPIs instead of one curve

To avoid picking a membrane “by eye,” the team assembles a multi-criteria evaluation framework (MCEF). It combines entropy weighting with TOPSIS. Five indicators go in:

caffeine removal,
flavour-marker retention index,
choline retention,
water permeance,
concentration increase.

Entropy weights which indicators most differentiate candidates. TOPSIS ranks membranes against ideal and anti-ideal points in that space. This is not a marketing “expert score.” It is a formal ranking on measurable KPIs.

In that ranking the best flat membrane is CGSM-1. The same line also looks promising in hollow-fibre form — a geometry closer to industrial modules than a lab sheet alone. That detail matters. A flat membrane proves channel chemistry. Hollow fibre asks about a scalable module shape.

Alginate, calcium and channels that quarrel

Sodium alginate enters between graphene oxide layers like a spacer. Calcium ions crosslink and stiffen the structure. The combined variant tries to take something from both. On paper it sounds like a cooking recipe: add a spacer, add a binder, done.

In nanochannels the recipe breaks. A spacer can widen the path for water and at the same time change what sticks to the walls. Crosslinking can improve selectivity toward caffeine and choke flow as a side effect. So the result does not form a simple ladder “unmodified < alginate < calcium < both.” It forms a compromise map that must be read KPI by KPI.

For a process engineer that matters more than a flashy press line. If a small chemistry change overturns the ranking, recipe quality control becomes part of the product. It is not enough to “have graphene.” You must repeatedly hit the same permeance–selectivity point.

What actually came out on coffee

In coffee filtration tests CGSM-1 lowered caffeine from 24.64 to 12.95 μg/mg. That is a 47.4% drop. At the same time measurable levels of HMF, trigonelline, NMP and choline — markers tied to flavour and matrix — were retained.

Read carefully. “Measurable levels” means the compounds did not vanish from the analytical setup. They do not automatically mean a sensory panel will call the brew equivalent to the starting coffee. The paper does not sell a consumer tasting. It sells process selectivity on an analytical set.

47% less caffeine is a strong lab result for a pass through a well-chosen GOM. It is still not a regulatory or shelf “caffeine-free” claim. EU and national thresholds for decaf coffee are stricter than “about half.” The authors themselves scope the claim to selective caffeine reduction in coffee liquor plus a practical membrane evaluation strategy — not a ready decaffeination line.

Reverse two: a lab win ≠ a factory win

The reverse closes without ornament. We have a membrane that wins the MCEF ranking on a five-KPI compromise. We have almost half the caffeine gone. We have retained markers. And we have a list of things the paper does not close.

There is no full mass balance at plant scale. No long-term stability under repeated CIP cleaning. No cost per square metre of membrane at beverage-typical lifetime. No sensory comparison with solvent decaffeination on the same bean lot. And there is no Polish lab protagonist — affiliations are Australian–Japanese.

For TRL09 that is interesting enough. It shows that a “coffee filter” at graphene nanochannel scale can be assessed like an engineered product: multi-criteria, with an honest trade-off, without a single success curve.

TRL without decoration

Today the work sits around TRL 3–4. There is material (modified GOMs), tests on coffee liquor, an MCEF ranking, a numeric result for CGSM-1 and a signal that hollow fibre makes sense. There is no pilot at a roastery, no sensory validation and no deployment path at an instant-coffee producer.

The risk is not only “does the membrane work.” The risk is process: fouling by oils and matrix fines, bean variability, regeneration, CGSM-1 batch repeatability, integration with an existing NF/RO line in the plant. Until those points sit outside the paper, an investor should read 47.4% as an experiment result, not a factory KPI.

Poland and Europe: a consumer angle, not an affiliation one

We do not glue on a Wrocław lab, because it is not in the affiliations. The PL/EU angle is market. Europe drinks a lot of coffee and has a dense processing and membrane-maker network. The question “can we cut caffeine without solvent and without killing flavour” hits a Warsaw shelf as much as Sydney.

The Polish market does not need its own paper to use the result. It has roasters, instant-coffee importers and a growing shelf of “gentler” products for people limiting caffeine. It also has chemical and membrane firms in the B2B chain that can read permeance KPIs the same way a Sydney lab does.

For local ingredient firms and instant-coffee plants the point is simple. If a graphene membrane keeps selectivity after hundreds of cycles and comes down in cost, it competes with process chemistry. If not — it remains a neat paper about TOPSIS on five KPIs. EU labelling rules for decaf coffee and rising demand for “milder” coffee still force the search for alternatives. This work delivers a compromise map, not a finished product.

Buyers, barriers and what is not in the paper

There are several buyers in this story. Instant-coffee producers that already run membrane stations. NF membrane suppliers looking for food niches beyond water and milk. Beverage-process integrators who can price operating energy against solvent.

Barriers are as concrete as the 12.95 μg/mg result. Show CGSM-1 lifetime under real fouling. Move the ranking from sheet to hollow-fibre module at scale. Close sensory and marker stability on variable feedstock. Finally compare cost per kilogram of reduced-caffeine coffee with existing methods.

The paper does not promise that. It shows that a well-chosen, modified graphene oxide membrane can cut almost half the caffeine in the lab and leave measurable matrix markers — and that membrane choice can rest on a formal MCEF, not one curve. That is enough to watch in TRL09. It is not enough to announce the end of classic decaffeination.

What remains of the reverse tomorrow

The opening reverse can now be said without a trick. Yes, GOM nanofiltration can, under controlled conditions, cut caffeine hard — here to about 47%. Yes, this is not “caffeine-free” coffee or a factory-ready line. Both are true, because they describe two result levels: selectivity in a test and the threshold of a commercial product.

A question the publication does not close remains. Will CGSM-1 keep its ranking after months of work? Will hollow fibre catch the sheet not only in the lab? Will a taste panel confirm what chromatography says about HMF, trigonelline, NMP and choline?

For now Sydney and Okayama show something more valuable than the slogan “coffee without chemistry.” They show a map of compromises and an honest method limit. In membrane foodtech, limits are currency. Without them marketing eats engineering.

Sources

Yihan Tian, Tongxi Lin, Yuta Nishina, Xiaojun Ren, Rakesh Kumar Joshi, Multi-criteria evaluation of graphene oxide nanofiltration membranes for decaffeination, Journal of Membrane Science (2026). DOI: 10.1016/j.memsci.2026.125992