Explore Phase Changes via AI-Assisted Simulations

← Back to AIMS

By Charles Xie ✉

Chemistry textbooks sort crystalline solids into four families — metallic, ionic, covalent, and molecular — and explain that the type of bond holding a crystal together decides how hard it is to take apart. It is usually left at that: a table of melting points to memorize. With AIMS, students investigate the connection instead. They talk to an AI agent, and it runs real molecular-dynamics experiments for them: it melts a crystal, grows one back from a seed, and brackets the melting point, then drops the measured data into the chat as a live chart. Nothing about melting or freezing is programmed in — the phase change emerges solely from the forces among atoms, ions, and molecules. We start with silicon, the crystal that matters most to today's industry: almost every computer chip, solar cell, and power transistor begins as a single crystal grown from a seed. In the window below, liquid silicon sits on a seed crystal only four atomic layers thick. Press Play and watch the liquid crystallize: atoms build tetrahedra onto the seed one layer at a time, each new layer copying the one below, until everything below the rough top layer is one single crystal. Further down is the same experiment for a metallic, an ionic, and a molecular crystal.

From Melt to Microchip

The chips in a phone, a laptop, or a data center start the same way: as a pure single crystal of silicon. Before a single transistor is patterned, silicon refined to electronic grade — better than 99.9999999% pure, fewer than one foreign atom in a billion — is melted in a quartz crucible at 1414 °C (1687 K). In the Czochralski process, a small seed crystal is dipped into the melt and pulled upward slowly, a few centimeters an hour, while seed and crucible rotate. The liquid freezes onto the seed atom by atom, copying its lattice, and grows into an ingot up to 300 mm across and as tall as a person, one continuous crystal from end to end. The ingot is then sliced into thin wafers and polished to a mirror finish, and only then do lithography, doping, and the hundreds of other steps of chip fabrication begin.

Why go to such lengths? Because a transistor depends on electrons moving through a perfectly regular lattice. Where two crystal grains meet, or where an atom sits out of place, the lattice is broken: electrons scatter, charges get trapped, and dopants spread unevenly, so devices built on a flawed crystal leak, slow down, or fail. Tens of billions of transistors on one chip need the same perfect crystal under every one of them. Growing a pure single crystal is therefore the first step in making a chip once the silicon is refined, and in many ways the most demanding.

The model above shows that first step at the scale where it actually happens — the solid–liquid interface, one atomic layer at a time. Students can see how the seed is copied: each arriving atom takes its place from the layer below, so the whole ingot inherits the seed's orientation. They can also see why the temperature at the interface must be held just right: a little below the melting point the crystal grows, a little above it, growth stops and the crystal can melt back. In the real process, the pull rate and the heat flowing through the interface are tuned to hold the interface just below the melting point for days, the whole time an ingot grows.

Why Start from a Seed?

Heat a crystal far enough and it always melts, but a liquid can be supercooled: cooled well below its freezing point without freezing. Unless dust or a container wall offers a template, a crystal must begin from a tiny ordered cluster that forms by chance, and most such clusters fall apart again before they can grow. This is nucleation. In a simulation the wait for a nucleus can run to nanoseconds or longer — far beyond what a browser can compute while a student watches.

The way around it is the same trick the semiconductor industry relies on to grow its silicon ingots: give the liquid a seed crystal to grow on, and the nucleation barrier disappears. AIMS does the same thing in miniature. Each of the four labs in this article holds a few hundred atoms: a crystal seed a few atomic layers thick at the bottom, its lowest layer or two anchored in place (drawn see-through) and the rest free to vibrate, under a freshly melted liquid with an open surface on top. Below the melting point the liquid freezes onto the seed, layer by layer. Above it the liquid stays liquid; at most a layer or two next to the seed take order from it. Right at the melting point, crystal and liquid are in balance and the front advances and retreats without getting anywhere. That makes seeded growth a direct way to find the melting point — and it works the same way for every kind of crystal.

Running Experiments by Asking

AIMS is an AI-native environment. A student does not need to know what a seeded-growth run is, how to pin a layer of atoms, or how to measure crystallinity. They simply ask — "can the liquid freeze again?", "what is the melting point of this model?", "what happens if I heat the crystal?" — and the built-in AI agent recognizes the intent, configures the lab, runs the experiment, and posts a live chart of the results in the conversation.

Four Types of Crystalline Solids

The table summarizes what the four labs show. Each "grows" and "stays liquid" temperature is a measured run, not a lookup; the model's melting point lies between them. "Time to freeze" is how long the melt takes to become crystal at the "grows" temperature, in simulated time (not the minutes you wait in a browser).

Covalent: silicon. In a diamond-structure crystal each atom is bonded to exactly four neighbors at the corners of a tetrahedron. An atom joining the crystal must find four partners at the right angles at once, so silicon grows slowly: about 60 picoseconds at 1500 K, ten minutes or more of watching in a browser. The topmost layer stays rough, because at this temperature a free silicon surface never quite settles. A run at 1600 K still grows, more slowly; at 1700 K the melt stays liquid, so the model's melting point comes out near 1650 K — close to the real 1687 K, as the potential's authors intended. That agreement matters, because this is what an ingot grower controls: how far below the melting point the crystal grows. Here students can see the same melt crystallize at one temperature and stay liquid at another.

Metallic: copper. In a metal, the outer electrons are shared by the whole crystal, and the bond between two atoms barely cares about direction. An atom arriving at the growth front only has to find a hollow among its neighbors, so copper freezes quickly: at 1100 K the melt is crystal in 20 to 30 picoseconds, two to three times faster than silicon. At 1300 K a layer or two next to the seed become ordered, but the rest stays liquid; at 1200 K the growth front advances and retreats without finishing. The model's melting point lies near 1200 K.

Ionic: rock salt. Sodium chloride sets a test the metal does not: two kinds of ions must each find the right place, every sodium surrounded by six chlorides and every chloride by six sodiums. An ion that lands next to one of its own kind is pushed out again. Even so, salt freezes about as fast as copper: 15 to 25 picoseconds at 1000 K. It only takes longer to watch, because the long reach of the electric force between ions makes each step more costly to compute. Watch the melt, too: molten salt is much less dense than the crystal, so the liquid starts out swollen and shrinks back as it freezes. At 1200 K the melt stays liquid, which puts the model's melting point between 1000 and 1200 K, bracketing the real value of 1074 K.

Molecular: nitrogen. Many solids around us — ice, sugar, wax, most drugs in a pharmacy — are built from molecules held together by forces far weaker than a chemical bond. Solid nitrogen is one of the simplest: N₂ dumbbells stacked on the same face-centered cubic lattice as copper, held there by van der Waals attraction so feeble that the crystal melts at 63 K. Here everything happens tens of kelvin above absolute zero instead of a thousand or more: the liquid freezes onto the seed at 35 K over 60 to 80 picoseconds, and stays liquid at 45 K. Look closely at the grown crystal: the molecular centers settle onto lattice sites, but the dumbbells keep tumbling. That is how many molecular crystals behave near their melting points — the positions freeze first, and the orientations lock in only much colder. (Real nitrogen tumbles on a hexagonal lattice; this model keeps the cubic one.) Because the molecules are held so weakly and the lab contains only a couple of hundred of them, chance plays a large part: the front advances in fits and starts, and now and then a run stalls halfway; a second run from Reset takes a different path. The Dry Ice Growth from a Seed example shows a harder case: carbon dioxide molecules must also line up in a fixed orientation, and they do.

Read down the "Model T_m" column and the lesson of the four families appears in measured numbers: the melting point follows the strength of what holds the crystal together, from a few tens of kelvin for molecules held by van der Waals forces to a thousand or more for metals, ionic crystals, and covalent networks. The "Time to freeze" column shows something textbooks rarely mention: how fast a crystal grows depends on more than how strongly it is bound. Silicon is slow because each arriving atom must find four neighbors at the right angles. Nitrogen is slow for a different reason: at 35 K its molecules move about four times more slowly than copper atoms do at 1100 K.

Models Make Predictions

None of the four simulations is told its melting point. Each uses a published interatomic potential — an embedded-atom potential for copper, Tosi–Fumi for salt, Stillinger–Weber for silicon, and a two-site potential for nitrogen. Most were fitted to properties of the cold crystal, such as its spacing, stiffness, and cohesive energy (for nitrogen, rescaled to match the last); Stillinger–Weber's authors also tuned it to silicon's melting point and liquid structure. For the others the melting point is a genuine prediction: salt lands close to experiment, while copper and nitrogen come out lower, likely because their potentials capture the solid better than the liquid. Students can ask the agent why, switch copper to a different potential with one sentence, and watch the crystal melt about 300 kelvin lower.

Other Ways to Study Melting

Watching the order collapse. Open one of the "Melting of" examples (copper, salt, silicon, or ice) and ask the agent to "show how the lattice order changes with temperature." It heats the crystal step by step and measures an order parameter that is high for a lattice and near zero for a liquid, choosing the measure that fits the lattice (rock salt and copper order differently). The curve stays high, then drops off a cliff. It is a vivid picture, but the cliff is not always the melting point: a crystal whose surface does not premelt, as salt's and silicon's do not, has nowhere for the liquid to start, so it can be superheated well past its melting point before it gives way. Salt holds on to 1250–1350 K, and silicon past 2000 K. Comparing the cliff with the seeded-growth bracket is a lesson in itself.

The heating curve and latent heat. Ask for "a heating curve" and the agent feeds energy into the crystal at a steady rate and plots temperature against heat added. As the lattice breaks up, energy goes into tearing particles out of their sites instead of making them move faster — the latent heat. In a small simulation the melting step sometimes appears as a flat plateau, and sometimes as a striking dip: the superheated solid collapses all at once, and the latent heat it suddenly absorbs is paid out of the particles' kinetic energy, so the temperature falls while heat keeps flowing in. For rock salt the gap between the solid and liquid branches gives a latent heat within a few percent of the handbook value.

Beyond melting. The same kind of conversation covers evaporation, too. In the Evaporation of Argon example the agent measures how the vapor pressure of a liquid rises with temperature and fits it with the Clausius–Clapeyron relation; in Evaporation of Water it opens the top of the box, pumps the vapor away, and records the pool cooling as its most energetic molecules escape.

Glass or Crystal: Annealing

Freezing onto a seed is not the only way a liquid can become a solid. Cool it fast enough and the atoms never find their lattice sites at all: they jam in a disordered pile, a glass. The Annealing of Amorphous Copper example (Metallic Crystals) opens on exactly that — 108 copper atoms melted at 3000 K and then quenched in half a picosecond. Orbit it and look along any direction: there are no rows. Press Play and, at 150 K, the atoms jiggle, but the pile stays a pile, stuck in a local energy minimum about 90 meV per atom above the crystal.

Annealing is how a glass escapes that minimum. Ask the agent "Anneal it: heat to 1000 K, hold for 10 ps, then cool back to 150 K over 20 ps." It drives the whole temperature schedule and charts the energy as it goes. During the hold — below the melting point, so nothing melts — the atoms gain just enough mobility to hop into place, the energy drops in a step, and rows appear. Then press Reset to get the glass back and try two other histories: melt it and quench it to 150 K in half a picosecond, and it is a glass again; heat it to 1200 K, where the glass melts, and cool it slowly to 150 K over 30 ps, and the liquid crystallizes on the way down. Both runs use the same atoms and end at the same temperature; the big difference is the cooling rate, and it decides whether a solid ends up a glass or a crystal. These runs are short — tens of picoseconds on 108 atoms — so each takes only a few minutes.

Annealing also brings the story back to chips. After a silicon wafer is cut from its ingot, dopant atoms are fired into it by ion implantation, which knocks the lattice near the surface out of order. The wafer is then annealed: heated so the damaged layer can rebuild itself on the perfect crystal underneath and the dopants can settle onto lattice sites. It is a solid-state cousin of seeded growth, and one of many times a chip passes through the furnace on its way to becoming a working device.

Why Use AI to Assist Scientific Simulations?

Experiments like these are among the hardest to set up by hand. A trustworthy melting point needs a seed, a melt, a thermostat, and a way to judge crystallinity — and each kind of crystal needs its own force field and its own measure of order. That is exactly the machinery a beginner does not yet know how to operate, and the AI agent operates it for them. Because it maps plain questions onto the right experiment, a student who wonders "why does salt melt at eight hundred degrees but nitrogen at minus two hundred?" can grow and melt both crystals and compare two measured charts, not just read a sentence.

And what comes back is measured, not assumed. A melting point is a single number in a table, but behind it are particles deciding, one collision at a time, whether to stay in a lattice or wander in a liquid. In AIMS the particles obey only the forces among them, and the crystalline fraction and the temperature are recorded as the simulation runs. When the copper melt freezes at 1100 K but stays liquid at 1300 K, students are seeing an emergent outcome of the bonding, not a drawing made to match a textbook.

Conclusion

This article uses phase changes to illustrate a new way to learn science: students converse with an AI agent that turns a plain-language question into a molecular-dynamics experiment and hands back measured data. The four families of crystals — metallic, ionic, covalent, and molecular — are no longer categories to memorize; students investigate how hot each must be to melt, how fast each can grow, and why. By pairing physically grounded simulations with an AI that can set them up and run them, AIMS turns a student's curiosity into a genuine scientific investigation of how matter changes its state.

← Back to AIMS