AI for scientific discovery

How does AI represent a physical object? Let's try to understand. In a simple way. A model is a pile of numbers.

It cannot hold a molecule. So the object has to become numbers first. That step is called representation. The obvious try.

Write down every atom. Carbon here, oxygen there, each with its position in space. Three numbers per atom. But now turn the molecule a little.

Every single coordinate changes. Same molecule, completely different numbers. So we want numbers that do not move when the object does. Describe the insides instead.

Distances and angles between atoms hold still when you rotate it. That is why chemistry uses a graph. Every atom is a node. Every bond is an edge joining two nodes.

Then each node carries a short list of numbers. Which element it is, how many neighbours it has, its charge. That list is a feature vector. One more trap.

Nothing says which atom is atom one. Renumber them and it is the same object. So the model passes messages between neighbours, not down a list. And whatever you leave out, no model can get back.

Drop the shape, and two molecules with the same atoms look identical. The representation sets the ceiling. Every field does this. A crystal becomes a repeating box of atoms.

Weather becomes a grid of cells. So, representation is choosing which numbers stand for the thing. Keep what matters. Stay the same when it turns.

Then the model can finally see it. Quick check now. One question is coming up. Let's see if it clicked.

What exactly is a protein folding puzzle? Let's try to understand. In a simple way. A protein starts out as a chain.

Beads on a string. Each bead is an amino acid, and there are only twenty kinds. But a flat line of letters does nothing. The chain folds up into one particular shape, and the shape is what does the job.

So why not try every fold? The chain bends at every joint. The number of possible shapes is larger than the number of atoms in the universe. So researchers stopped asking for the whole shape.

They asked something smaller instead. For any two beads, how far apart do they end up? Ask that for every pair, and the chain turns into a grid. One box for bead five against bead ninety.

Filling in the boxes is the puzzle. Each box is now a single small guess. Near, or far. And the boxes constrain each other, so one wrong answer fights with its neighbours.

But where do the clues come from? From other species. The same protein, lined up across thousands of them, stacked into rows. Here is the trick.

Two positions that always change together are probably touching. A mutation in one only survives if its partner shifts as well. So every pair that moves together fills in a box. Thousands of hints, scattered over the grid, squeezing the empty boxes into place.

So the puzzle is a grid of pairs, and evolution hands you the first clues. Solving that grid comes next. Quick check now. One question is coming up.

Let's see if it clicked.

How does protein shape prediction work? Let's try to understand. In a simple way. A grid of distances is not a shape yet.

What you want is three numbers for every bead. Where it sits in space. So here is how the model starts. Every bead stacked at the same point.

A ball of nothing. Then it pushes them apart. Each bead is more than a dot. It carries a tiny frame.

A position, and the direction it faces. The model predicts a turn and a shift for it. And it moves all of them at once. Nothing is simulated.

No atoms bouncing around, no waiting for the chain to settle. How does it know where to push? Two tables sit behind it. The rows of related species, and the grid of pairs.

Numbers flow between them. The grid also obeys geometry. If the first bead is near the second, and the second is near the third, then the first cannot be far from the third. Checking every triangle forces the guesses to agree.

Then the neat part. The shape it just produced is fed straight back in as input. Same model, another pass. Three more times.

Each new pass gets to see its own last answer. A loop that came out wrong gets straightened, and the fold tightens. And it learned this from real structures. Around two hundred thousand of them, measured in labs over fifty years.

So the shape comes out as coordinates, guessed in one shot, then refined by feeding it back. Not simulated. Predicted. Quick check now.

One question is coming up. Let's see if it clicked.

What exactly is a confidence score? Let's try to understand. In a simple way. A folding model always returns a shape.

Even for a protein it cannot fold, it draws a full answer. The picture never looks unsure. So the model prints a second thing beside the shape. A score for every bead, zero to one hundred.

One score per position, not one for the whole protein. A packed core can sit at ninety while a dangling tail sits at thirty. But scoring against what? Nobody knows the true shape yet.

The model is predicting how wrong it expects to be. It learned that from proteins whose shape was already known. Predict, measure the real error at each bead, train a second output to guess it. The number reads as a scale.

Above ninety, trust it down to the atoms. Around seventy, only the backbone. Below fifty, do not trust the shape. A low score is not always failure.

Many of those stretches have no fixed shape in the cell. The score doubles as a floppiness detector. Bead scores still miss something. Two pieces can each be right and still sit wrongly against each other.

So a second score covers pairs: expected error in one bead if you line up on the other. One warning. The confidence is itself a prediction, from the same model. Usually right, not always, and never checked against an experiment.

So the score is the model's own estimate of its error, bead by bead. Read it before you read the shape. Quick check now. One question is coming up.

Let's see if it clicked.

How does drug design from a predicted shape work? Let's try to understand. In a simple way. Most medicines are small molecules.

A tiny thing that sticks to one protein and jams it. Sticking is physical. It has to fit the surface. The predicted shape hands you exactly that.

A dent in the surface. Its walls, its depth, and the chemical groups lining it. That is the pocket. A molecule binds when its bumps land in the pocket's hollows, and its charges sit opposite the charges on the wall.

Shape and chemistry at once. So here is what the computer does. Take one candidate molecule, drop it into the pocket, and try thousands of positions and twists. That is docking.

Each pose gets a number. How much surface touches, how many hydrogen bonds form, whether opposite charges line up. The best pose is the molecule's score. Now run that over a catalogue.

Hundreds of millions of molecules, scored and ranked on a cluster in days. Nobody touched a test tube yet. What comes out is not a medicine. It is a shortlist.

A few thousand worth ordering and testing at the bench. Newer models skip the catalogue. Hand them the pocket, and they draw a molecule to match it, atom by atom. Two catches.

The pocket walls are the shakiest part of a predicted shape. A good score only guesses that it sticks. Sticking is not curing. So the shape gives you a pocket, docking ranks millions against it, and the bench tests the top few.

Quick check now. One question is coming up. Let's see if it clicked.

What exactly is a weather grid? Let's try to understand. In a simple way. Weather is measured in a messy way.

Stations on land, buoys at sea, balloons twice a day, satellites overhead. None of it lines up. But a model cannot read a pile of scattered readings. It needs the same numbers, in the same places, every time.

So we lay a grid over the planet. Every box in that grid holds a few numbers. Temperature. Pressure.

Wind, split in two directions. Humidity. And the grid is not flat. It is stacked upward through thirteen levels of the atmosphere.

So the planet becomes a cube of numbers. How fine is it? A quarter of a degree is common. Boxes roughly twenty five kilometres wide.

About a million on every level. Now the awkward part. Most of those boxes were never measured. Nobody has a thermometer in the middle of the Pacific, six kilometres up.

So the gaps are not invented. You start from the last forecast, the model's guess for right now, and nudge it toward the readings that did arrive. That is data assimilation. Run that for every hour, back across decades, and you get one consistent record.

Forty years, all on the same grid. And to an AI model, that cube is just a stack of images. Each variable is one channel, the way a photo has three. Same shape every hour.

So the grid is not a picture of the weather. It is the dataset. Scattered readings in, a tidy cube out. Quick check now.

One question is coming up. Let's see if it clicked.

How does an AI weather forecast work? Let's try to understand. In a simple way. The old way is physics, done properly.

Air is a fluid, and we know the equations it obeys. A supercomputer marches the grid forward, one small step at a time. But the steps have to be tiny. A few minutes of weather per step, or the numbers blow up.

Ten days is thousands of steps. The AI model skips all of that. It never solves an equation. It is handed the grid as it looks now, and it predicts the grid six hours later.

It learned that jump from the record. Forty years of past grids, cut into pairs. The world now, the world six hours on. Millions of examples.

So one step for the AI is six hours wide. The same six hours a solver crawls through in a hundred tiny steps. For a longer forecast, its own answer goes back in as the next input. Six hours, twelve, eighteen.

Forty steps gets you ten days out. That is where the speed comes from. An hour on a supercomputer becomes under a minute on one chip. The work did not vanish.

It moved. Training costs weeks of compute, once. Every forecast after that is cheap. One catch.

It only ever saw the weather the physics record contained. Push it far out and the storms it draws start to blur. So it is not faster physics. It is a learned shortcut.

Same grid in, same grid out, none of the steps between. Quick check now. One question is coming up. Let's see if it clicked.

What exactly is extreme weather detection? Let's try to understand. In a simple way. Picture forty years of weather, one grid of numbers every hour.

That is billions of boxes. Nobody scrolls through that by eye. The interesting part is tiny. Hurricanes, heat waves, atmospheric rivers.

To study them, you first have to find every single one. The old way is a hand written rule. Wind above this speed, pressure below that value. It works, until two teams pick different numbers and count different storms.

Because an event is not one box crossing one number. It is a shape across many boxes. A spiral of wind. A ribbon of wet air.

So experts mark examples by hand. They draw around the storm. These boxes are inside the event, these are outside. A few thousand of those.

Then the model reads the grid like an image and labels every box. Storm, or not. The output is a mask. That task is called segmentation.

Here is the trap. These events are rare. Fewer than one box in a thousand sits inside one. Always answering no scores ninety nine percent.

So accuracy is the wrong score. The rare boxes get weighted up in training, and the model is judged on overlap with the marked event. Then it sweeps the whole archive. Forty years scanned in hours.

Every event comes back with a date, a place, and a track. So it is not a threshold. It is a shape, learned from marked examples, then swept across decades of grids. Quick check now.

One question is coming up. Let's see if it clicked.

What exactly is the space of possible materials? Let's try to understand. In a simple way. A material is not just its ingredients.

It is which elements, and how the atoms are stacked. Carbon one way is graphite. Stacked differently, it is diamond. So count the options.

Three elements from about a hundred usable ones is already over a hundred thousand trios. Add ratios and packings, and you reach billions. A century of lab work has made and measured a couple of hundred thousand. The rest has never been touched.

And you cannot make them all. One compound can take weeks in a furnace. So the map gets drawn on a computer. The trick is substitution.

Take a crystal structure that already works, and swap new elements into the same positions. A few thousand templates give millions of candidates. Most swaps are nonsense, so each candidate is scored. A physics calculation estimates its energy, which says whether those atoms would really sit that way.

The test is the convex hull. A candidate survives only if nothing else built from the same atoms sits lower in energy. Otherwise it falls apart into that instead. So the map is a catalogue.

Each entry stored with its elements, its arrangement, and its distance above the hull. And it keeps growing. One recent search added two point two million predicted crystals. About four hundred thousand looked stable.

So the universe of materials is not a place you visit. It is a list you generate, then filter down to what could hold together. Quick check now. One question is coming up.

Let's see if it clicked.

How does predicting a material's properties work? Let's try to understand. In a simple way. You have a candidate crystal on paper.

Every atom in its place. The question is what it will do. Conduct? Bend?

Survive heat? Two honest answers exist. Make it and measure it. Or solve the electron physics on a computer.

Weeks of work, or hours of computer time. So we learn a shortcut. Take crystals whose numbers we already know, and fit a model that reads the atoms and predicts the number. The first move turns the crystal into a graph.

Every atom becomes a node. Every pair close enough gets an edge, carrying their distance. Each node starts as a short list of numbers standing for its element. Iron gets one list, oxygen another.

Learned during training. Then the graph talks to itself. Every atom collects its neighbors' lists, mixes them with its own, and updates. That is one round of message passing.

Run three or four rounds. After the second, an atom has heard from its neighbors' neighbors. So each atom describes its surroundings, a few bonds wide. Now average every atom list into one list for the whole crystal.

A small final layer turns that into one number. The band gap, in electron volts. That whole pass takes milliseconds, not hours. So you screen a million candidates overnight, then run the real physics on the best hundred.

So the model never solves any physics. It recognizes neighborhoods it has seen before. It ranks candidates well, and proves nothing. Quick check now.

One question is coming up. Let's see if it clicked.

How does an AI write a recipe for a new material? Let's try to understand. In a simple way. A search gives you a formula and an arrangement of atoms.

It does not tell you how to make it. Nobody has ever heated that compound. A recipe is specific. Which powders you start with, what temperature you fire them at, for how long, and in what gas.

Traditionally a chemist guesses that from experience. Each guess is a day in the furnace, and most come out wrong. So the training data was dug out of the literature. Software read the methods sections of millions of papers and pulled out thousands of real recipes.

Now the model reads a target formula and proposes the starting powders. They must supply every element, and their extras must leave as gas. Carbonates are favorites. A second part predicts the numbers.

The firing temperature, often around nine hundred degrees. The hold time. And the gas: air, or argon. The output is not one recipe.

It is a ranked handful, five or so, ordered by the model's confidence. And confidence here means familiarity. The model copies recipes that worked for chemically similar targets. Unusual chemistry, with no relatives in the papers, gets weak suggestions.

So the recipes go to a robot lab. It mixes, heats, and X-rays the product to see what formed. Failures come back as new training data. So the model never reasons about the reaction.

It generalizes from what has worked before. It proposes, and the furnace decides. Quick check now. One question is coming up.

Let's see if it clicked.

How does training on simulated experiments work? Let's try to understand. In a simple way. A model needs examples with the answer attached.

In science that answer is a measurement, and measurements are slow. One sample, one week at a bench. But there is a second source of answers. The equations.

A computer can solve the physics for a structure nobody has ever made. So the simulation becomes the label. Feed in the atoms, run the calculation, keep the number. That pair is one training example.

The win is volume. Nobody has to be there, and a thousand runs go at once on a cluster. Millions of labeled examples. Now look at what the model really learns.

Not nature. It learns to imitate a program, in milliseconds instead of hours. So it inherits that program's mistakes. If the calculation reads low for every metal, the model copies the same offset on every prediction.

That is the reality gap. Simulations are clean. Real samples carry impurities, defects, and a messy instrument. The model saw only the clean version.

So you buy a little reality. A few hundred real measurements, and the model is trained on the difference between simulation and bench. You also choose what to simulate next. Let the model mark where it is least certain, run those calculations, and add them in.

So simulated training buys labels by the million, with one condition. The model can only be as right as the simulator it copied. Quick check now. One question is coming up.

Let's see if it clicked.

How does AI pattern discovery work? Let's try to understand. In a simple way. Here is how a person hunts for a pattern.

Plot one column against another and look at the picture. Two things at a time. But one modern experiment records two hundred columns per sample. That is twenty thousand pairs of plots.

Nobody is flipping through those. The model never takes two columns. It takes the whole row at once, all two hundred numbers, for every sample in the archive. And that is where a missed pattern hides.

In a combination. Column four matters only when column ninety is low. Neither plot alone shows it. Give it no labels.

It places every sample as a point in a space. Similar ones land together, and a clump appears that nobody named. Here is the real asymmetry. A person looks where the theory says to look.

The model has no opinion about which column is supposed to matter. But a score is not a finding. So you change one column, hold the rest, and watch how far the prediction moves. That ranks the columns.

Now the trap. It will happily key on the scanner brand instead of the disease, because in your data those two travel together. And a pattern is still a correlation inside one dataset. The model has no notion of cause.

What you have is a lead worth testing. So the advantage is not insight. It is every column at once, every record, and nothing it expects to find. Quick check now.

One question is coming up. Let's see if it clicked.

What exactly is an anomaly flag? Let's try to understand. In a simple way. An anomaly detector never learns what a discovery looks like.

It learns what normal looks like, from millions of ordinary records. The score is one number. How badly this record fits everything seen before. Nothing in it says new physics.

Only, this does not fit. And two very different things score high. A signal nobody has recorded before. And a detector channel that died on Tuesday.

Both are just unusual. In a real archive, glitches outnumber discoveries by thousands to one. So the top of the ranked list is almost entirely broken data. Normal means whatever was in the training set.

Point a second telescope at the same sky, and its ordinary quirks get flagged as strange. Some of it is your own processing. A step that patches dead pixels leaves a smooth blob the model never saw. So it lights up.

The first test is boring. Line the flagged events against temperature, calibration runs, the clock. Real physics does not care what time it was. A fault does.

The second test is repetition. A different instrument, a different run. An artifact stays inside its own hardware. A real signal comes back.

One more possibility. The thing is real, and already published. It was missing from the training set, so the model met it as a stranger. So an anomaly flag is a work order, not a result.

Rule out the hardware, the pipeline, the literature. What survives is worth a look. Quick check now. One question is coming up.

Let's see if it clicked.

How does validating a surprising AI hypothesis work? Let's try to understand. In a simple way. The model hands you a sentence.

This compound should melt near two hundred degrees. That is a claim. Nothing has been measured yet. And you cannot check it against the archive it came from.

Ask that archive and it agrees. It has to. The pattern was pulled out of it. So the first move is to turn the claim into one number.

Not, this looks promising. A specific value, for a case nobody has run yet. The cheapest test is data the model never saw. An old experiment sitting in a drawer.

Another lab's records. If the number misses there, stop. Stronger still, write the prediction down before the experiment exists. Then run it.

Now nobody can quietly move the target afterwards. The model only saw two things travel together. The lab changes one of them on purpose and watches the other move. Correlation cannot fake that.

And surprise has a price. A result that fights fifty years of chemistry needs a taller pile of evidence than one that agrees with it. So you repeat it. Second lab, different instrument, fresh batch.

If the effect only appears on your machine, it was your machine. Then ask why it happens. A confirmed number with no mechanism is a fact you cannot extend. A mechanism tells you what else to try.

Most candidates die somewhere in there, and that is the job. The model narrows the search. The experiment decides. Quick check now.

One question is coming up. Let's see if it clicked.