Computers now invent materials
AI models can propose thousands of new crystals for batteries, solar cells or catalysts, faster than anyone can make them.
A careful planner for making new materials
From a material on a screen to a recipe a lab can follow.
Mat2Bot works out how a crystal could be made: which chemicals, how hot, for how long. It shows the evidence behind every step, and it says plainly how sure it is, including when the honest answer is "we don't know yet".
Why a planner like this matters
Each one still needs a recipe
Which chemicals, in what amounts, at what temperature, in what order? A material on a screen is not a material in a jar.
Confident and wrong is the worst outcome
A bad recipe wastes weeks of lab work. So every answer comes with its evidence, its doubts, and an honest level of certainty.
How it works
How Mat2Bot thinks, in seven steps
Scroll through the steps; the picture changes as you go.
Pin down exactly what to make
A material is more than a formula. The same atoms can stack in different patterns, and each pattern is a different material. Zinc oxide is usually "wurtzite", but squeezed hard enough it becomes a denser "rock-salt" form. Mat2Bot starts from measured crystal structures, so it knows exactly which one you mean.
Try both forms of zinc oxide in the picture.
Gather the evidence
It collects what is already known: published experiments, measured data tables and chemistry databases. Every number it uses keeps a link to its source. A checking program re-reads the exact sentence in the paper, so nothing is taken on faith.
When nobody has made the material yet, it looks at how similar materials were made: their starting chemicals, how hot they were fired and in which gas. From these established methods, each shown on its own, it drafts a recipe, clearly marked as a hypothesis to test.
Check the chemistry
Then it runs the science: how much will dissolve, what will form, whether the material is stable. Hot water holds more salt than cold water. Cool it, and the extra salt has nowhere to go: its ions lock into a crystal.
Cool the beaker and watch the crystal grow. If a model is asked about conditions it was never built for, Mat2Bot says so instead of guessing.
Hunt for reasons it could fail
A good scientist asks what could go wrong. For each kind of synthesis Mat2Bot works through a checklist of specific doubts: could a different crystal form appear? Is a data table being stretched too far? Each doubt ends up ruled out, still open, or confirmed.
Say how sure it is
A fixed rulebook, not a gut feeling, turns the evidence and the doubts into one of four answers: well supported, supported with a caveat, plausible but not yet proven, or rejected. It never claims more than the evidence allows.
Write the recipe, check the lab
The plan becomes a step-by-step recipe any lab could follow. It is then matched to a real lab: which steps a robot can do, which need a person, what is missing. Hazards such as heating a flammable liquid are kept for a person, and nothing runs without a named person's sign-off.
Let an independent auditor check it all
Everything is sealed in a package: the evidence, the calculations, the decisions and a tamper-evident log. A separate auditor program, which never calls the planner, opens the package and recomputes the results from scratch. Every case below also has a trail, "How Mat2Bot got there": the same records as a map you can click through, stage by stage, down to each number's source.
Same atoms, different packing.
Published experiment (McBride et al. 2003)
…to 101 °C and maintaining it at that temperature for periods of up to 8 hours…quote found in the paper
Measured data (NIST ThermoML)KCl in saturated water at 25 °C: 0.2648 g per g of solutionvalue re-read from the database file
Crystal structure (COD 2300450)zinc oxide, wurtzite: cell edge 3.25 Åmeasured structure, open database
Every number keeps its source.
60 °C
potassium chloride ions
Doubts checked for calcite (calcium carbonate), test case 4.
RejectedPlausibleCaveatWell supported
Who does each step in the robot lab, for test case 3 (crystals with ethanol).
The auditor's checks on test case 4, rerun when this site was built.
Explore
Twelve test cases, in plain language
Each case was chosen to test one ability, from easy salt crystals to a material that needs enormous pressure, and two new materials proposed by MEIDNet, an AI model that designs crystals. Open one to see what Mat2Bot concluded, how sure it is, the recipe it wrote, and the full technical record.
These results were computed in advance. To plan a material of your own, type it into the box on the front page; the engine's code stays private.
Can you trust it?
It was graded against predictions written in advance
For every test case we wrote down, before the code it tests, what Mat2Bot ought to conclude: 195 separate predictions. Then we built it and checked every one. The misses are published with their reasons, never tuned away.
- 94 problems were found by four independent reviews before release, and all were addressed.
- Every plan comes as a sealed package that a separate auditor recomputes.
- In a blind test on 500 known oxides, its draft starting chemicals matched those of the published recipes about 3 times in 4; a simpler rule did better, and that is published too.
- Every number says where it comes from: a quoted sentence, a database entry, a calculation by a named rule, or a stated choice.
Words you will meet