AI Water Discovery Explained: Two Liquids in One Glass, for Kids

Pour a glass of water and you are looking at one liquid. Physicists are not so sure. Here’s the AI water discovery explained in one sentence: in July 2026, a University of Osaka team trained a neural network to judge which of 16 mathematical “descriptions” of a water molecule’s neighborhood best tells apart water’s two competing liquid forms, and the winner separated them with near-perfect accuracy. That sentence hides three ideas a 10-year-old can handle: water has two personalities, scientists had 16 rival ways of spotting them, and a computer program acted as the referee. The rest of this article takes those ideas apart, then puts them on your kitchen table.

Key Takeaways

• The paper (Yoshikawa, Shikata, Kim, and Matubayasi, Communications Chemistry, July 2026) did not discover a new kind of water. It ranked 16 existing ways of measuring water’s structure using a neural network as an impartial judge.
• The winning descriptor, the Local Structure Index (LSI), told cold-water snapshots from warm-water snapshots with an AUC of at least 0.991 in one set of simulations. Several popular descriptors did far worse.
• The data was entirely simulated: 1,000 virtual water molecules, modeled with a standard recipe called TIP4P/2005, between 200 and 300 kelvin. No real water was measured.
• The two-liquid idea itself goes back to a 1992 Nature paper and got direct experimental support in 2020. This AI paper is a tool-sharpening step, not the final proof.
• The teachable lesson for kids is about method: AI in science often works by grading human hypotheses, and the “grades” are only as good as the questions humans wrote.

What the Osaka team actually did

The study is a machine-learning comparison of structural descriptors for supercooled water, published as “Machine learning evaluation of structural descriptors for supercooled water” in Communications Chemistry 9(1) on July 8, 2026, by Kohei Yoshikawa, Kokoro Shikata, Kang Kim, and Nobuyuki Matubayasi of the University of Osaka’s Division of Chemical Engineering (DOI 10.1038/s42004-026-02097-1; open preprint on arXiv).

A “structural descriptor” is a single number you can compute for one water molecule that summarizes how its neighbors are arranged. Is it sitting in a neat, open, tetrahedral cage like the one in ice? Or is it squeezed into a crowded, disordered cluster? Over the past three decades, physicists proposed at least 16 different formulas for that number. Each has fans. Nobody had run a fair contest.

So the Osaka group built one. In the university’s own words, the neural network “used what it had learned to compare how 16 descriptors differentiated between LDL and HDL structures at different temperatures,” according to senior author Matubayasi in the EurekAlert release. LDL and HDL are the two suspected liquid forms: low-density liquid and high-density liquid.

The AI water discovery explained: what the network computed, step by step

This is the part most news coverage skipped, and it’s the part your kid should hear first.

Step 1: Make the data. The team ran molecular dynamics simulations in GROMACS using the TIP4P/2005 water model, a well-tested set of equations that treats each water molecule as a small rigid object with charges on it. They simulated 1,000 molecules at temperatures from 200 K to 300 K (about −73 °C to 27 °C), under two conditions: fixed density (1 g/cm³) and fixed pressure (1 bar). Every few time steps they saved a “snapshot” of where all 1,000 molecules were.

Step 2: Compute each descriptor. For every snapshot, they calculated all 16 descriptors for all 1,000 molecules. That gives 16 different lists of 1,000 numbers per snapshot.

Step 3: Ask the network a simple question. The neural network’s job was binary classification: given a list of 1,000 descriptor values from one snapshot, is this a cold snapshot or a warm one? The network was small: 1,000 input nodes, two hidden layers, a sigmoid output that spits out a probability. It was trained with a standard loss function (binary cross-entropy) and a standard optimizer (Adam).

Step 4: Score each descriptor by how easy it made the question. If the network can tell cold from warm almost perfectly using descriptor A but only slightly better than a coin flip using descriptor B, then A captures the temperature-driven structural change and B doesn’t. The score they used is AUC, the area under the ROC curve, where 1.0 is perfect and 0.5 is guessing.

Step 5: Open the black box. They applied LIME (Local Interpretable Model-Agnostic Explanations), which fits a simple linear model around each prediction to show which input values pushed the answer toward “cold” or “warm.” That let them see which molecules the network was paying attention to, not just whether it got the answer right.

What it got right

The Local Structure Index (LSI) achieved “nearly perfect discrimination, with AUC ≥ 0.991 under isochoric conditions,” per the preprint. Three others also scored well: ζ (zeta, which measures the gap between a molecule’s hydrogen-bonded and non-bonded neighbors), NTCHB (a network-communicability measure counting hydrogen bonds), and the plain hydrogen-bond count N_HB.

What it got wrong, or couldn’t say

Three honest caveats sit in the paper itself. First, descriptors like Q₄, Q₆, and coordination number were sensitive to an arbitrary cutoff distance; change the cutoff and the ranking shifts. Second, every descriptor got worse between 260 K and 300 K, because at warmer temperatures density fluctuations blur the two structures together. Third, the network classified temperature, not LDL versus HDL directly. Temperature is a proxy. A descriptor that tracks temperature well is probably tracking the liquid-liquid change, but “probably” is doing some work there. And all of it lives inside one water model, TIP4P/2005; whether the ranking transfers to other models, or to real water, is unstated.

The analogy: grading 16 ways to describe a crowd

Imagine a school gym full of kids. Sometimes they are spread out in a calm, organized pattern, each kid at arm’s length from four friends. Sometimes they crowd together in messy clumps. You can’t see the whole gym; you only get one number per kid.

Sixteen teachers each propose a number: “count how many kids are within one step,” “measure the distance to the fifth-closest kid,” “check whether the four nearest kids form a diamond shape.” Which number best reveals whether the gym is in calm mode or crowd mode?

The Osaka approach: show a computer thousands of gyms, tell it which were calm and which were crowded, and let it try to guess using only one teacher’s number at a time. The teacher whose number makes guessing easiest wins. That’s the whole trick. The AI did not invent a new number. It refereed.

Why water has two personalities in the first place

Liquid water is odd. It is densest at about 4 °C, not at its freezing point, which is why ice floats and lakes freeze from the top down. Its compressibility and heat capacity climb sharply as you cool it below 0 °C without letting it freeze. In 1992, Poole, Sciortino, Essmann, and Stanley proposed in Nature that these anomalies come from a hidden transition between two liquid forms, ending at a “liquid-liquid critical point” deep in the supercooled region.

That idea sat as a hypothesis for nearly 30 years. In November 2020, Kim and colleagues published in Science an experiment using X-ray laser pulses to watch high-density liquid water convert to low-density domains on timescales of 20 nanoseconds to 3 microseconds, well before it crystallized (3 to 50 microseconds). That separation of timescales is the strongest experimental evidence so far that the two liquids are real.

The 2026 Osaka paper doesn’t add evidence for the transition. It sharpens the ruler used to measure it. That distinction matters when you explain it to a kid: “the AI proved water is two liquids” is wrong; “the AI figured out which ruler measures the difference best” is right.

How to Teach Your Kid About the AI Water Discovery

Ages 5–8: The Ice Cube Float Test

Fill a clear glass with water and drop in an ice cube. Ask: does it sink or float? Then say the strange part out loud: almost everything else gets heavier when it freezes and sinks. Water is the weird one. Freeze a small sealed bottle of water overnight (leave room at the top) and show how the ice bulged. That bulge is the “open cage” structure the Osaka team’s descriptors are trying to detect. No AI talk needed yet; the goal is for them to remember that water is unusual.

Ages 9–12: Build a Descriptor by Hand

Spread 30 coins on a tray. Arrange them once in a neat honeycomb pattern and once in messy clumps. Have your kid invent a “descriptor”: one number per coin, like “how many coins are touching this one.” Compute it for five coins in each arrangement. Ask which number was more different between the neat and messy trays. Now explain that the Osaka team had 16 such numbers, and instead of doing it for 5 coins they did it for 1,000 molecules across thousands of snapshots, and used a computer to score which number worked best.

Ages 13+: Train a Tiny Classifier

Open a spreadsheet. Column A: 40 rows of “neighbor count” for the neat pattern (mostly 6s with a few 5s). Column B: 40 rows for the messy pattern (random values from 2 to 8). Have your teen pick a threshold that best separates the columns and count the errors. Then repeat with a worse descriptor (say, coin diameter, which is the same everywhere). This is a one-feature classifier and an ROC curve in miniature. If they want to go further, Google’s Teachable Machine lets them train an image classifier on photos of the two trays in ten minutes and see a confidence score.

The question to ask: “If the computer got the answer right 99% of the time using the LSI number, does that prove water has two liquid forms, or does it prove something smaller?”

The 16 descriptors, ranked: what the network learned

The full paper reports every descriptor’s AUC across both conditions. This table groups them by how the classifier performed and by what each one is trying to capture.

Descriptor (paper’s symbol)	What it measures, in plain terms	How the network rated it
LSI (Local Structure Index)	How “gappy” the spacing is between a molecule’s nearby neighbors	Best: AUC ≥ 0.991 at fixed density
ζ (zeta)	Distance gap between hydrogen-bonded and non-bonded neighbors	Strong, near the top
NTCHB	How well-connected a molecule is through hydrogen bonds	Strong
N_HB	Simple count of hydrogen bonds	Strong, surprisingly competitive
q_tet, q_n	How close the four nearest neighbors are to a perfect tetrahedron	Useful, but weaker at warm temperatures
d₅	Distance to the fifth-nearest neighbor	Moderate
V₄, V₅, V₄₋₅	Interaction energies with the 4th and 5th neighbors	Moderate
Ψ (psi), NTC	Topological and network measures	Moderate
Q₄, Q₆, C.N., V_voro	Bond-orientational order, coordination number, Voronoi volume	Sensitive to arbitrary cutoff choices; weaker

Read the third column as “how easy this number made the cold/warm guess,” not as a measure of scientific truth.

What to actually do at home

Separate “discovered” from “measured”

When your kid brings home a headline like “AI cracks water’s mystery,” ask them to say what was measured. Here the honest answer is “which formula best tracks a structural change in a simulation.” Practicing that translation is the single most useful science-literacy habit for the next decade, and this paper is a clean, low-stakes place to start.

Do one physical experiment before any screen

The float test and the bottle-bulge test take ten minutes and cost nothing. Kids remember what their hands did. The neural-network story lands better after they’ve held the evidence that water expands when it freezes. If your family enjoys it, the physical-versus-chemical change guide is a natural next step.

Let them invent a bad descriptor on purpose

Ask your 9-to-12-year-old to propose a number that would not tell neat from messy (coin color, coin size). Then ask why. Understanding why a measurement fails is what the Osaka paper’s cutoff-sensitivity caveat is really about.

Try the supercooling demo as a follow-up

Once the “two liquids” idea has settled, the supercooled-water bottle trick shows what “below freezing but still liquid” looks like. It’s covered step by step in our companion piece on supercooled water and neural networks in the kitchen.

What not to do

Don’t tell your kid the computer “understood” water. It didn’t. It found a boundary between two labeled piles of numbers. Anthropomorphizing the network makes the real method harder to see later, and the real method is the part that transfers to every other AI-in-science story they’ll meet.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can restate the result as “a program scored 16 formulas” rather than “AI discovered two waters.” If they still say “discovered,” rerun the coin-tray activity.
• Month 2 red flags: They treat “the AI said so” as the end of an argument in other contexts (homework, news). Ask what the AI was trained on and what question it was asked, every time.
• Month 3 self-check: Can they explain why ice floats using the words “open structure” or “cage”? Can they name one thing the Osaka study could not prove? If yes to both, the concept stuck.

Frequently Asked Questions

Did AI actually discover that water is two liquids?

No. The two-liquid hypothesis dates to a 1992 Nature paper by Poole and colleagues and received direct experimental support in a 2020 Science paper. The 2026 Osaka study used a neural network to rank 16 existing structural descriptors by how well they track the change in simulations. It sharpened a measurement tool; it did not make the original discovery.

Is this something my kid can see in a glass of water?

Not directly. The two liquid forms show up clearly only in supercooled water, below 0 °C, where the differences become pronounced. What kids can see is the consequence: ice floating, sealed bottles bulging in the freezer, and lakes freezing from the top. Those everyday effects trace back to the same open molecular structure the descriptors measure.

What is a “descriptor,” in words a 10-year-old would use?

A descriptor is one number you calculate for one molecule that describes its neighborhood, like “how many friends are within arm’s reach” or “how diamond-shaped is the group around it.” The Osaka team tested 16 of them. The best one, LSI, measures how uneven the gaps are between a molecule’s nearest neighbors.

Why did the researchers use simulated water instead of real water?

Because you cannot photograph individual molecules in liquid water fast enough to compute descriptors. Molecular dynamics simulations, here using the TIP4P/2005 model and 1,000 molecules, give exact positions at every moment. The tradeoff is that results are only as realistic as the model, which the authors flag as a limitation.

Does this change anything about drinking water or weather?

Not this year. Better descriptors help researchers study ice formation in clouds, protein folding in cells, and how water behaves under pressure, all of which depend on local structure. The practical payoff is upstream: better tools for the scientists working on those problems, not a change to your tap.

---

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 of the same paper with full methods. arXiv 2605.00415. https://arxiv.org/abs/2605.00415
3. The University of Osaka. (2026, July 6). “Out of order: using AI to decode the bizarre personality of water.” EurekAlert! https://www.eurekalert.org/news-releases/1134535
4. ScienceDaily. (2026, July 8). “Scientists used AI to crack one of water’s biggest mysteries.” https://www.sciencedaily.com/releases/2026/07/260707025047.htm
5. 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
6. 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
7. 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