The Phase Transition Hiding in Your Cacio e Pepe

A Phase Diagram for Dinner

Cacio e pepe should be the easiest dish in the Roman repertoire. Cook pasta. Reserve some of the starchy water. Melt Pecorino Romano into that water off the heat, whisk in black pepper, toss with pasta, serve. Three ingredients, one pan, no oven. And it fails so reliably that Italian cooking forums have entire threads devoted to nothing else: you follow the steps, and instead of a glossy cream the cheese seizes into rubbery threads and grit, sitting in a thin, separated liquid. Cooks call it "the sauce broke." Restaurant kitchens that serve it at volume have worked out folk fixes over generations — more starch, less heat, add the cheese slower — without anyone writing down why any of it works.

In January 2025, eight physicists posted a preprint that treats that failure as exactly what a condensed-matter physicist would call it: a phase transition in a multi-component mixture, with a real phase diagram, a real critical point, and a controllable order parameter. It ran in Physics of Fluids in April, authored by Giacomo Bartolucci, Daniel Maria Busiello, Matteo Ciarchi, Alberto Corticelli, Ivan Di Terlizzi, Fabrizio Olmeda, Davide Revignas and Vincenzo Maria Schimmenti [1]. Several of them do this kind of physics for a living on considerably less delicious systems — phase separation in polymer mixtures and the physics of biomolecular condensates inside living cells [14] — and the paper reads like people who noticed their day job sitting in their dinner. Cooking has drawn serious physicists' attention before, from pasta's own fracture mechanics [2] to a broader review of fluid dynamics across the kitchen [3]; this paper's contribution is to bring the same seriousness specifically to a sauce.

What Actually Goes Wrong in the Bowl

Pecorino Romano is not a simple solid that melts. It is an emulsion holding two populations of milk protein in suspension: casein, which sits in the cheese as large, roughly spherical clusters called micelles, and whey proteins, smaller globular proteins folded into a compact shape that keeps their sticky interior hidden. When the cheese is intact and cold, both populations mostly keep to themselves. The whole system holds together as one continuous, workable mass.

Grate that cheese into hot water and both populations start to change. Whey proteins denature — the folded shape opens up, exposing the hydrophobic patches that used to be tucked inside — and once denatured, they are unusually eager to stick to themselves, to each other, and to the outside of casein micelles [5][6]. The casein micelles, meanwhile, aggregate on their own as they warm [7]. Push enough heat into the system and what was individually-dispersed protein starts finding itself more attracted to more protein than to the water around it. Aggregates form, then grow, then merge into visible clumps big enough to trap fat and leave behind separated liquid. That is what a home cook watches happen in real time when the sauce "breaks," and it is what the paper's authors, tongue only partly in cheek, name the Mozzarella Phase: a texture that would be exactly right in a different dish, and a disaster in this one.

The Backward Physics of a Broken Sauce

Here is the part of the paper that overturns a piece of kitchen folklore. Home cooks who have had a sauce break tend to blame it on the sauce cooling unevenly, or on the cheese hitting cold pasta, and the general advice is to work off the heat and to keep things from getting too cold too fast. The measured physics runs the other way. The paper reports that with no starch present, cheese mixed with hot water forms large clumps once the mixture crosses about 65 °C — the clumping is triggered by getting too hot, not by cooling down [1].

The formal name for this shape of phase behavior is a lower critical solution temperature, or LCST. Most people's intuition for mixing runs the opposite way: oil and vinegar separate when they're cold and blend when you shake them warm and vigorously; salt dissolves better in hot water than cold. An LCST system is the reverse of that intuition. Below the critical temperature the components mix into one homogeneous phase; above it, the mixture becomes thermodynamically unhappy being one phase and splits into two, one protein-rich and one protein-poor, in exactly the pattern the authors observed. Whey proteins denaturing on heating supplies the switch: cold, undenatured whey protein stays folded and sociable toward water; hot, denatured whey protein flips to a state that prefers sticking to other protein instead [1][5]. Push a cacio e pepe past that switch and the physics does not care how carefully you built the emulsion up to that point.

Turning a Ruined Sauce Into Data

Knowing the direction of the effect is not the same as mapping it. To build an actual phase diagram, the team needed a way to measure "how clumped" a sample was, reproducibly, across many combinations of ingredients and temperatures — something more rigorous than a cook's eyeball.

Their method: mix cheese, water, and a chosen starch concentration, spread the mixture in a thin layer across a transparent Petri dish, and backlight it from below. Aggregated protein blocks light; a smooth sauce lets it through evenly. Ramp the temperature from 50 °C to 95 °C in 5 °C steps, over roughly 40 minutes per run, photographing the dish at every step. Then run the photographs through an image-processing pipeline: normalize to greyscale, smooth with a Gaussian filter (σ = 1 pixel) to suppress noise, binarize with a quantile threshold swept between 0.05 and 0.25 to isolate the dark aggregate regions, clean the result up with watershed segmentation, small-object removal, dilation and hole-filling, and finally approximate every detected aggregate as an ellipse. The major axis length of that ellipse becomes the order parameter — the single number, growing from zero in a smooth sample to the width of the dish in a fully clumped one, that turns "does it look broken" into something you can plot [1]. Every temperature ramp was run at least twice.

The Line at One Percent

With a working order parameter, the team ran two systematic sweeps. First: hold cheese and water at equal mass (a 1:1 ratio) and vary starch concentration across 0%, 0.5%, 1%, 1.5%, 2%, 2.5%, 3% and 4% of the cheese's mass, at every temperature step from 50 to 95 °C. Second: fix starch at 1% and vary the cheese-to-water ratio instead, to see how concentration on its own moves the boundary [1].

The result is the paper's central claim, and it is closer to a bright line than a fuzzy trend. Below 1% starch (relative to cheese mass), the sauce enters the Mozzarella Phase — system-wide clumping — once heated past its critical temperature. At or above 1%, the same cheese and water stay smooth all the way up to 95 °C. Push starch too far the other direction, past roughly 4%, and a new failure mode appears: not clumping, but a sauce so thick and gluey with gelatinized starch that it stops being pleasant to eat. The paper's practical sweet spot sits at 2–3% starch, comfortably inside the smooth region and short of the stiff one [1].

The Free Energy Underneath the Boundary

A phase boundary is not just an empirical line; the paper grounds it in the physics of polymer solutions. Mix a solute made of large molecules (proteins, starch) into a solvent (water), and the classic way to write the free energy of that mixture per unit volume is the Flory–Huggins form, built independently by Paul Flory and Maurice Huggins in 1942 [10][11]:

Here $\phi$ is the protein mass fraction, $n$ captures how large the protein-bearing species is relative to the solvent, and $\chi$ is the interaction parameter — how much the solute and solvent would rather not mix. The first two terms are entropy: any mixing at all is favored just because there are more ways to arrange a mixed system than an unmixed one. The last term is energy, and it is the one that can win. Raise $\chi$ high enough and the free energy curve develops two separate minima instead of one, meaning the system is happier splitting into a protein-rich phase and a protein-poor phase than staying uniform. That split is exactly what the paper measures as clumping.

What makes this an LCST system rather than the more familiar kind is that heating increases $\chi$ here, rather than decreasing it. The paper attributes that rise to precisely the biochemistry above: warming denatures whey protein and drives casein micelles to aggregate, both of which make protein-protein contact more energetically favorable relative to protein-water contact as temperature climbs [1]. The paper also lets $n$, the relative size term, grow with temperature, which they interpret as evidence that starch is not just diluting the protein but forming larger starch–protein complexes as the system heats [1][9].

Where the Line Sits: the Binodal and Its Minimum

Running the second sweep — fixed 1% starch, varying cheese-to-water ratio — lets the authors trace the actual phase boundary (the binodal) in terms of protein concentration rather than starch concentration. Plotted as protein mass fraction $\phi$ against temperature, the boundary comes out as a parabola opening upward, with its minimum — the easiest point at which to trigger phase separation — sitting at a temperature slightly above 60 °C and a protein mass fraction of $\phi \approx 0.134$, which is what you get from cheese and water combined at a plain 1:1 ratio by mass [1]. To turn a cheese quantity into a protein fraction at all, the authors used the actual nutritional label of Pecorino Romano DOP: 27 g of protein per 100 g of cheese [1] — so at 1:1 cheese to water, half the mixture's mass is cheese, and $0.5 \times 0.27 \approx 0.134$, which is exactly the minimum the binodal traces out.

A Sharper Fix: Trisodium Citrate

Starch is not the only additive the paper tests. Trisodium citrate — the same salt used to keep processed cheese smooth, and a common trick in the "modernist" fondue and cheese-sauce literature [4] — gets the same treatment: swept across concentrations relative to cheese mass and run through the identical temperature ramps.

The comparison is instructive. Starch defends the sauce gradually: as its concentration rises through the danger zone, clump size falls off over a spread of a percent or two before disappearing. Citrate does something closer to a switch. At 1% citrate the Mozzarella Phase still forms at high enough temperature; at 2% the sauce is fully stabilized, and 3% confirms it — a transition the authors describe as sharp, in explicit contrast to starch's gradual one [1]. The likely reason is mechanistic rather than incidental: citrate is a small ion that can chelate the calcium holding casein micelles together and directly disrupt protein-protein bonding, while starch's protection is closer to a steric and complexation effect — getting physically in the way of aggregating proteins and possibly binding some of the denatured whey directly [1][9]. A direct chemical disruption of the bond should show up as a sharper phase boundary than a physical crowding effect, and that is what the data show. The tradeoff: citrate blunts the cheese's flavor more than starch does, which is presumably why nobody's grandmother reaches for it.

From Phase Diagram to Recipe

Having a phase diagram is only useful if it produces a protocol, and the paper closes with exactly that: a recipe built directly on the measured boundary rather than on tradition. Below is the paper's own recommended procedure and quantities, for two servings [1].

Notice what this recipe is not built on. It is not built on a chef's palate, or on how a dish "usually" turns out, or on doing it the way an Italian grandmother does it, because Italian grandmothers disagree with each other about method constantly. It is built on a measured boundary in a phase diagram, engineered to sit inside the smooth region with room to spare. That is the entire argument of the paper in one recipe card.

The Club's Arithmetic: Checking the Recipe Against the Diagram

The paper states its threshold as a percentage and hands over a recipe in grams. Nobody actually measures cheese sauce as a percentage while standing at a stove, so we did the conversion ourselves — entirely our own arithmetic, run on the paper's own reported numbers, and none of it appears in the paper itself.

Where the recipe sits on the starch axis. The recipe calls for 5 g of starch against 200 g of cheese:

That lands squarely inside the paper's own "ideal" 2–3% window, and it is 2.5× the bare 1% survival threshold — not a knife-edge margin, but not an enormous one either. Cut the starch by more than half relative to the recipe and you are back below the clumping line.

Where that is in a measuring spoon. Nobody keeps a gram scale by the stove for 5 g of starch. Using the standard kitchen conversion of roughly 8 g of cornstarch per tablespoon — so about 2.67 g per teaspoon, a common kitchen reference figure and not a number from this paper at all:

Call it a scant two teaspoons. It is worth sitting with how small that is: the difference between a smooth cacio e pepe and the Mozzarella Phase, in this recipe, is roughly the difference between two level teaspoons of starch and one.

Where the recipe sits on the protein axis. This is the calculation we think matters most, because it connects the recipe directly back to Figure 3. The binodal's minimum — the single easiest point in the whole diagram at which to trigger phase separation — sits at φ≈0.134, measured for cheese and water alone with no starch present. Using the same definition the paper uses for φ, and the same 27 g-protein-per-100 g figure for Pecorino Romano [1], the recipe's own cheese and water (200 g cheese, 150 g water, starch excluded from this axis exactly as the paper excludes it) give:

The actual dish, in other words, sits at a protein concentration close to — and just past — the single most clump-prone point on the entire phase diagram. That is not a coincidence you'd want to bet against: a home cook using less water than the recipe calls for pushes φ even higher, deeper into the danger zone, while more water pulls it toward either edge of the parabola where the two-phase region is narrower. It is also, we think, the real answer to why the 1% starch line matters as much as the paper's data say it does: the traditional dish is not made at some safely dilute protein concentration where clumping is hard to trigger. It is made almost exactly where clumping is easiest.

A consistency check. The paper separately describes its citrate alternative as "2.5% by cheese mass," and its stated quantity is the same 5 g against the same 200 g of cheese [1]. That is exactly the 2.5% our own division above produces for the starch line too — a small thing, but it is reassuring when arithmetic done two different ways on two different ingredients lands on the same number.

Why This Is Actually Legitimate Physics

It would be easy to read all of this as a joke dressed up in equations — physicists cannot resist measuring things, so they measured dinner. We think that reading misses what is genuinely being demonstrated here. A phase transition does not care what industry supplies its ingredients. Casein micelles aggregating under heat, whey proteins denaturing and turning sticky, a polymer additive (starch) sterically stabilizing a colloid against that aggregation, an LCST separating one phase from two — every piece of that vocabulary belongs equally to a dairy science paper, a materials science paper on protein gels, or a kitchen. The authors are not stretching physics to cover a novelty topic. They are pointing out that the physics was already there, unexamined, because nobody had bothered to run the experiment on a system this mundane.

There is a second reason the choice of system is not incidental. Several of the paper's authors work, in their non-culinary research, directly on liquid–liquid phase separation and biomolecular condensates — the same physics that governs how cells organize their interior without membranes, currently one of the more active areas of biophysics [12][13][14]. Giacomo Bartolucci co-authored a 2023 PNAS paper on cyclic phase separation as a mechanism for selecting molecular sequences during the origin of life [14], work that lives in exactly the same theoretical framework — free energy landscapes, interaction parameters, binodal curves — used here to explain a broken sauce. That is not a coincidence of authorship; it is the actual point. The mathematics of Flory and Huggins, built in 1942 to describe rubber and polymer melts, turns out to organize a startlingly wide range of real systems, from the protein droplets inside a living cell down to a pot on a stove, and demonstrating that range on a system every reader has personally ruined is, if anything, better pedagogy than doing it on a system nobody outside a lab has ever touched.

None of that erases the humor. "Mozzarella Phase" is a deliberately funny name, chosen by people clearly enjoying themselves. Funny and rigorous are not opposites here; the joke is precisely that the rigor was warranted all along.

What We Would Want to See Next

A few open threads, in the order we would chase them.

Other cheeses. The whole phase diagram rests on Pecorino Romano's specific protein content and its particular ratio of casein to whey. Parmigiano Reggiano, aged differently and with a different moisture content, would almost certainly move the boundary; nothing in the paper as described here tests that.

Other starches. Potato and corn starch are named as interchangeable, but they gelatinize at different temperatures and to different final viscosities. Whether the 1% line holds identically for both, or drifts, is a natural next measurement.

The mechanism behind the size parameter. The paper infers that starch-protein complexes grow with temperature from how the Flory–Huggins size parameter $n$ behaves in the fit, rather than from a direct structural measurement. Confirming that with, say, light scattering on the actual mixture would turn an inference into a demonstrated mechanism.

Whether the margin generalizes. Our φ≈0.154 calculation is specific to the exact recipe given, at Pecorino Romano's exact protein content. A recipe that used less cheese relative to water, or a different hard cheese with a different protein fraction, would sit at a different point on the same parabola — possibly a safer one, possibly not. The interesting general question is whether traditional recipes across cuisines that emulsify hard cheese into hot liquid tend to cluster near the most demanding point of their own phase diagrams, the way this one apparently does, or whether that is peculiar to cacio e pepe.

References

1. 1.

   Bartolucci, G., Busiello, D. M., Ciarchi, M., Corticelli, A., Di Terlizzi, I., Olmeda, F., Revignas, D. & Schimmenti, V. M. (2025). Phase behavior of Cacio e Pepe sauce. Physics of Fluids 37, 044122. doi:10.1063/5.0255841

2. 2.

   Ball, P. (2010). Pasta physics. Nature Materials 9, 539. doi:10.1038/nmat2793

3. 3.

   Mathijssen, A. J. T. M., Lisicki, M., Prakash, V. N. & Mossige, E. J. L. (2023). Culinary fluid mechanics and other currents in food science. Reviews of Modern Physics 95, 025004. doi:10.1103/RevModPhys.95.025004

4. 4.

   Bertsch, P., Savorani, L. & Fischer, P. (2019). Rheology of Swiss Cheese Fondue. ACS Omega 4, 1103-1109. doi:10.1021/acsomega.8b02424

5. 5.

   Anema, S. G. & Li, Y. (2003). Association of denatured whey proteins with casein micelles in heated reconstituted skim milk and its effect on casein micelle size. Journal of Dairy Research 70, 73-83. doi:10.1017/S0022029902005903

6. 6.

   Anema, S. G. (2021). Heat-induced changes in caseins and casein micelles, including interactions with denatured whey proteins. International Dairy Journal 122, 105136. doi:10.1016/j.idairyj.2021.105136

7. 7.

   Nicolai, T. & Chassenieux, C. (2021). Heat-induced gelation of casein micelles. Food Hydrocolloids 118, 106755. doi:10.1016/j.foodhyd.2021.106755

8. 8.

   Liu, Y., Tan, Z., Xu, X., Liu, J., Zhu, B. & Dong, X. (2023). The formation of rice starch/casein complex by hydrothermal treatment and its use in stabilizing high internal phase emulsions. Food Hydrocolloids 144, 108995. doi:10.1016/j.foodhyd.2023.108995

9. 9.

   Jenkins, P. J. & Donald, A. M. (1998). Gelatinisation of starch: a combined SAXS/WAXS/DSC and SANS study. Carbohydrate Research 308, 133-147. doi:10.1016/S0008-6215(98)00079-2

10. 10.

    Flory, P. J. (1942). Thermodynamics of High Polymer Solutions. The Journal of Chemical Physics 10, 51-61. doi:10.1063/1.1723621

11. 11.

    Huggins, M. L. (1942). Thermodynamic properties of solutions of long-chain compounds. Annals of the New York Academy of Sciences 43, 1-32. doi:10.1111/j.1749-6632.1942.tb47940.x

12. 12.

    Brangwynne, C. P., Eckmann, C. R., Courson, D. S., Rybarska, A., Hoege, C., Gharakhani, J., Jülicher, F. & Hyman, A. A. (2009). Germline P granules are liquid droplets that localize by controlled dissolution/condensation. Science 324, 1729-1732. doi:10.1126/science.1172046

13. 13.

    Hyman, A. A., Weber, C. A. & Jülicher, F. (2014). Liquid-Liquid Phase Separation in Biology. Annual Review of Cell and Developmental Biology 30, 39-58. doi:10.1146/annurev-cellbio-100913-013325

14. 14.

    Bartolucci, G., Calaça Serrão, A., Schwintek, P., Kühnlein, A., Rana, Y., Janto, P., Hofer, D., Mast, C. B., Braun, D. & Weber, C. A. (2023). Sequence self-selection by cyclic phase separation. Proceedings of the National Academy of Sciences 120, e2218876120. doi:10.1073/pnas.2218876120