Supercooled Water Experiment Kids Can Run, Plus the AI Behind It

The supercooled water experiment kids ask for again and again needs three things: a sealed bottle of purified water, a freezer, and patience. Leave the bottle alone for roughly two hours and the water cools below 0 °C without freezing. Tap it, or pour it over an ice cube, and it turns to slush in about a second. That’s the whole demo, and it works because ice needs a starting point to grow from. It also happens to be a doorway into a July 2026 paper from the University of Osaka, where a neural network was used to judge which mathematical descriptions of water’s molecular structure best track what happens in this exact below-zero, still-liquid state.

Key Takeaways

• Supercooled water is liquid water below 0 °C. It stays liquid because ice crystals need a “nucleation site” (dust, a scratch, an existing ice crystal) to begin forming, and purified water in a smooth sealed bottle offers none.
• A home freezer typically gets water to somewhere between −5 °C and −15 °C before it freezes on its own; in the lab, pure water can be pushed to roughly −40 °C, the homogeneous nucleation limit (Debenedetti, 2003).
• The slush you see, rather than solid ice, is a latent-heat effect: freezing releases heat, which warms the remaining water back toward 0 °C and stops the freeze partway.
• The Osaka study (Yoshikawa et al., Communications Chemistry, July 2026) simulated water from 200 K to 300 K, a range that includes your freezer, and found that the Local Structure Index separates cold from warm structures with AUC ≥ 0.991.
• The experiment fails often, and that’s the point. Kids learn more from a bottle that froze early than from one that worked the first time.

What supercooled water is, and why it sits still

Supercooled water is liquid water held below its normal freezing point of 0 °C without crystallizing. It is not a trick or an unstable fluke. It’s what water does whenever nothing gives it a reason to freeze.

Ice is a crystal. Crystals grow from a seed: a speck of dust, a microscopic scratch on the container wall, a tiny existing ice crystal. Give water a seed and it freezes at 0 °C, as expected. Take away every seed, and the molecules keep jostling around in liquid form well below zero, waiting. The University of Illinois Physics Van puts it plainly: water “can get ‘stuck’ in its liquid state as it cools off, even below its freezing point,” and what it needs to get unstuck is “a little piece of dust or other impurity in the water, or even a scratch on the bottle” (Physics Van, University of Illinois).

You see this in the sky more often than you’d think. Freezing rain is supercooled rain. The National Weather Service describes drops that cool below freezing while falling and then, “when the supercooled drops strike the frozen ground (power lines, or tree branches), they instantly freeze, forming a thin film of ice” (NWS winter weather training). An ice storm is your bottle experiment, scaled up to a county.

The supercooled water experiment kids can run in one evening

You need 4 to 6 unopened bottles of purified or distilled water (plastic, not glass), a freezer with some clear space, a bowl, and a few ice cubes.

1. Lay the bottles flat in the freezer, not touching each other or the walls. Set a timer for two hours.
2. At two hours, open the freezer gently and look without touching. If none have frozen, close it and check every 15 minutes.
3. The moment one bottle shows ice, the rest are probably supercooled. Take one out slowly, holding it by the cap.
4. Option A: tap the bottle hard on the counter. Watch a wave of ice crystals travel from the point of impact.
5. Option B: open the cap and pour the water slowly onto an ice cube in the bowl. It stacks up as slush.

The procedure above follows the version published by Home Science Tools, which also names the two common failure modes: leaving the bottles too long (they freeze solid) and pulling all the bottles out at once (they warm up before you use them).

Expect failure. Tap water, with its dissolved minerals and particles, often freezes at 0 °C like it’s supposed to. Bottles with scuffs or dents nucleate early. A freezer that runs cold (−18 °C is typical) can push water past what it can hold. None of this is a problem; it’s the data.

What’s happening at the molecular level: the Osaka connection

Here is where the kitchen meets the paper. In July 2026, Kohei Yoshikawa, Kokoro Shikata, Kang Kim, and Nobuyuki Matubayasi of the University of Osaka published “Machine learning evaluation of structural descriptors for supercooled water” in Communications Chemistry (DOI 10.1038/s42004-026-02097-1; preprint). Their subject was not the bottle. It was the question of how to measure what water is doing in that below-zero liquid state.

What they computed. They simulated 1,000 water molecules using the TIP4P/2005 model across temperatures from 200 K to 300 K (−73 °C to 27 °C). Your freezer bottle, at maybe 260 K, sits inside that range. For each simulation snapshot, they calculated 16 different “descriptors,” each a single number per molecule summarizing how its neighbors are arranged. Then they trained a small neural network to answer one yes/no question: given the descriptor values from a snapshot, is this cold water or warm water?

What data they used. Only simulated data. The neural network never saw a real water molecule, which the authors state as a limitation. But the simulation model they chose, TIP4P/2005, is a widely used one that reproduces water’s density maximum near 4 °C and other anomalies.

What it got right. The Local Structure Index (LSI), which measures how uneven the gaps are between a molecule’s nearest neighbors, gave the network “nearly perfect discrimination, with AUC ≥ 0.991 under isochoric conditions.” Three others also did well: ζ (zeta), a hydrogen-bond-network measure called NTCHB, and the plain hydrogen-bond count.

What it couldn’t do. Every descriptor got weaker between 260 K and 300 K, meaning right around the temperatures of your freezer and fridge, the cold-versus-warm difference is blurrier. And the network classified temperature, not the two liquid forms directly; temperature is a stand-in. The paper sharpens a ruler. It doesn’t prove the two-liquid picture, which rests on a 1992 Nature proposal by Poole and colleagues and a 2020 Science experiment by Kim and colleagues that watched high-density water convert to low-density domains in nanoseconds.

The link to the bottle, stated carefully. In the simulations, as temperature drops, more molecules adopt the open, ice-like, tetrahedral arrangement the descriptors are built to detect. The supercooled water in your bottle is in that regime. It is structurally primed. When a nucleation site appears, the ordered cage forms fast because much of the local geometry is already close to ice. That’s why the freeze-front races through the bottle instead of creeping.

If you want the full breakdown of the paper, including the analogy of grading 16 teachers’ rulers, see our companion piece on the AI water discovery explained for kids.

Temperature by temperature: what water is doing and what your kid can see

Temperature	What the water is doing	What your kid can observe
100 °C (373 K)	Boiling; molecules escape as vapor	Steam from a kettle
4 °C (277 K)	Densest point; anomaly caused by open structure competing with compact structure	Not visible at home, but explains why lake bottoms stay at 4 °C
0 °C (273 K)	Normal freezing point; ice and water coexist if a seed exists	Ice cubes in a glass of water
−5 to −15 °C (258–268 K)	Supercooled liquid, if pure and undisturbed	The bottle demo; tap it and watch slush form
−18 °C (255 K)	Typical home freezer setting	Bottles left too long freeze solid
~ −40 °C (~231 K)	Homogeneous nucleation limit; water freezes without any seed (Debenedetti, 2003)	Lab only; cloud droplets can reach this before freezing
200–300 K (−73 °C to 27 °C)	Osaka simulation range	The full span the neural network was trained on

The middle rows are your experiment. The bottom rows are why physicists cared enough to build a neural network to study this regime.

How to Teach Your Kid About Supercooled Water

Ages 5–8: The Tap-to-Freeze Show

You handle the freezer and the bottle; your kid does the tap. Before the tap, ask them to predict: “Is this bottle colder or warmer than an ice cube?” (It’s colder than 0 °C, which surprises most kids.) After the slush forms, hand them a cube from the freezer and let them feel that the “frozen” bottle is now slushy, not rock-hard. Vocabulary to introduce: “seed.” Ice needs a seed to start, and the tap made one.

Ages 9–12: The Nucleation Race

Freeze three bottles side by side: one tap water, one distilled water, one distilled water with a pinch of salt dissolved in it. Have your kid predict, in writing, which will freeze first and which will supercool longest. Check every 15 minutes and log the results. Tap water and salty water usually behave differently from pure distilled water, but not always, and the “not always” is the lesson. Repeat on a second night. Ask which variable they’d change next.

Ages 13+: The Temperature Log and the Yes/No Question

Put a cheap digital thermometer probe into one bottle (through a small hole in the cap, sealed with tape). Log the reading every 10 minutes and plot it. Watch for the moment after nucleation when the reading jumps back up toward 0 °C; that’s latent heat, and it’s why you get slush. Then pose the Osaka team’s question: if you had only one number per molecule, which number would tell you this water is in its cold state? Have them read the preprint’s descriptor list and pick two they can explain in one sentence each.

The question to ask: “The bottle was below zero and still liquid. What was missing, and why did one tap supply it?”

What to actually do at home

Buy purified water, not “spring” water

Spring and mineral waters carry dissolved solids that act as nucleation sites. Distilled or purified (reverse-osmosis) water in unscratched plastic gives you the best odds. Four bottles for a few dollars is the whole budget.

Run it twice, one week apart

The first run teaches the phenomenon. The second run, with your kid choosing one thing to change, teaches the method. That two-step pattern is the same one the Osaka group used: first establish the effect exists, then build a fair test to measure it.

Keep a failure log

Write down every bottle that froze early and why you think it did. Scuffed? Touching the freezer wall? Left too long? A page of failures is more scientifically honest than a single success, and it mirrors the paper’s own limitations section.

Connect it to weather the next time it rains near freezing

Freezing rain is the same physics on a landscape. When a winter forecast mentions “freezing rain” or “ice storm,” bring up the bottle. Kids who have felt supercooled water in their hands understand why a road can be wet one minute and glass the next. Our rain gauge and weather station project pairs well with this.

What not to do

Don’t use glass bottles. Supercooled water can expand fast on freezing and glass can crack. Don’t let young kids handle bottles straight from the freezer for long; a −10 °C plastic bottle is uncomfortably cold on small hands. And don’t tell them the neural network “discovered” anything about the bottle. It ranked measuring tools using simulated water. The honest version is more interesting anyway.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can explain why the bottle didn’t freeze using the word “seed” or “nucleation,” and can name one reason a bottle might freeze early.
• Month 2 red flags: They repeat “the AI figured out water” without being able to say what data the network trained on (simulations) or what question it answered (cold or warm?). Reopen the paper’s method section together.
• Month 3 self-check: Ask them to design a third version of the experiment with a variable you haven’t tried (bottle size, freezer position, water brand). If they can state a prediction and a reason, the method stuck, not just the trick.

Frequently Asked Questions

Is it safe to drink the supercooled water afterward?

Yes. Supercooling changes nothing chemically; it’s the same water at a lower temperature. The Illinois Physics Van notes the phenomenon occurs naturally and poses no danger. The only real hazard is cold plastic on small hands and the small chance of a cracked container if you use glass. Stick with plastic bottles.

How cold does the water actually get in a home freezer?

Usually somewhere between −5 °C and −15 °C before it nucleates on its own, depending on purity and how still the bottle stays. A typical freezer runs at −18 °C, so if you leave the bottles long enough they will freeze solid regardless. Pure water in a lab can be supercooled to roughly −40 °C, the homogeneous nucleation limit.

Why does it turn to slush instead of solid ice?

Freezing releases latent heat. As the first crystals form, they warm the surrounding water back toward 0 °C, and at 0 °C the remaining water has no reason to freeze. The result is a mix of ice crystals and liquid, which looks like slush. Solid ice would need the freezer to remove that heat again over time.

What does the Osaka neural network have to do with my bottle?

The Osaka team simulated water across 200 to 300 kelvin, which includes freezer temperatures, and used a neural network to rank 16 ways of describing molecular structure in that regime. Your bottle is a real-world sample of the supercooled state they studied. The paper does not analyze bottles; it improves the tools scientists use to measure what’s happening inside them.

My bottles froze solid every time. What went wrong?

Most likely one of three things: the water had dissolved minerals (use distilled), the bottles were scratched or dented, or they stayed in too long. Try two hours, then 15-minute checks. Keep bottles flat and away from freezer walls and fans, and don’t bump the freezer while it runs.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Yoshikawa, K., Shikata, K., Kim, K., & Matubayasi, N. (2026). “Machine learning evaluation of structural descriptors for supercooled water.” Communications Chemistry, 9(1). https://www.nature.com/articles/s42004-026-02097-1
2. Yoshikawa, K., et al. (2026). Preprint with full methods. arXiv 2605.00415. https://arxiv.org/abs/2605.00415
3. Debenedetti, P. G. (2003). “Supercooled and glassy water.” Journal of Physics: Condensed Matter, 15(45), R1669–R1726. https://iopscience.iop.org/article/10.1088/0953-8984/15/45/R01
4. University of Illinois Physics Van. “Supercooled Water in the Freezer.” https://van.physics.illinois.edu/ask/listing/1618
5. National Weather Service. “Winter Weather” training page on freezing rain and supercooled drops. https://www.weather.gov/source/zhu/ZHU_Training_Page/winter_stuff/winter_wx/winter_wx.html
6. Poole, P. H., Sciortino, F., Essmann, U., & Stanley, H. E. (1992). “Phase behaviour of metastable water.” Nature, 360, 324–328. https://www.nature.com/articles/360324a0
7. Kim, K. H., et al. (2020). “Experimental observation of the liquid-liquid transition in bulk supercooled water under pressure.” Science, 370(6519), 978–982. https://www.science.org/doi/10.1126/science.abb9385
8. Home Science Tools. “Super-cooled Water Experiment.” https://learning-center.homesciencetools.com/article/super-cooled-water-science-project/