When the shortlist comes back empty
Sometimes no existing material sits where you need it in the property space. The creator proposes candidates that might — as hypotheses with a route to test them, clearly labelled as what they are.
From composition to a testable claim
- 01 Property gap Where the shortlist does not reach
- 02 Composition Stoichiometry, with each element's role
- 03 Structure Candidate lattice and why it holds
- 04 Synthesis A first attempt you can actually run
What a concept contains
| Composition | A proposed stoichiometry, with the role each element plays in the target property. |
|---|---|
| Candidate structure | The crystal structure or microstructure the concept assumes, and why that symmetry supports the property. |
| Structure–property rationale | The physical argument: electronegativity and ionic radii, valence electron count, Hume-Rothery rules where they apply. |
| Synthesis route | A concrete first attempt — route, precursors, atmosphere, temperature window — so the idea is testable. |
| Failure modes | Where the concept is most likely to fall over, and what to measure first to find out. |
Creative, but bounded by physics
The interesting failure mode of generative models in materials is confident nonsense: a stoichiometry that violates valence, a structure that could not tolerate the ionic radii involved, a synthesis above the precursor's decomposition temperature.
So the creator argues from the quantities a materials scientist would use — electronegativity difference for bond character, radius ratios and tolerance factors for whether a substitution fits, valence electron concentration for intermetallics, Hume-Rothery rules for solubility. Where the argument is weak it has to say so.
Treat the output as an experiment plan
Take the composition to structure search first and see whether the chemical system is already populated — a hull distance near zero for a nearby phase tells you a lot about what will actually come out of the furnace. Use the selector before any of this: inventing is only the right move once the shortlist genuinely cannot reach your targets.
Generated materials are unverified proposals. Nothing here has been synthesised or characterised. Validate in the lab before any of it reaches a design.
- Crystal systems — symmetry constraints
- Periodic table — radii and electronegativity
- Material classes — what bonding permits
Questions about the creator
Is the output a real material?
It is a hypothesis. The creator proposes compositions that are physically plausible given the chemistry, not compounds verified to exist. The synthesis route is there precisely so the proposal can be tested rather than believed.
How is plausibility enforced?
The reasoning is grounded in real chemistry — electronegativity differences, ionic radii and tolerance factors, valence electron counts, Hume-Rothery solubility rules — and the model is instructed to be honest about uncertainty rather than confident by default. You can then check whether anything similar already exists using structure search.
Who is this actually for?
Researchers scoping a direction before committing lab time, and engineers who have exhausted the shortlist and need to know whether the gap in the property space is fundamental or merely unexplored.
Can I check whether the composition is already known?
Yes — search the chemical system in the structure search, which queries the Materials Project, OQMD and the Open Materials Database. If DFT structures already exist for that system you will see them, along with formation energy and hull stability where the provider publishes them.
Put an engineer's AI on the problem
Free plan, no card required. 20 AI requests a month, and the whole knowledge base stays open to everyone.