Why this page is short
Because the honest description of what Creator does is short, and the interesting things to say about generative materials design are cautionary rather than promotional.
A generative model conditioned on a target property will hand you compositions. Some will be sensible. Some will already exist. Some will be unmakeable. The model cannot distinguish between those cases, and neither can the number it attaches to its output.
What “E1” means here in practice
An E1 value requires the model name and version, the training set, the quantity predicted and the reported uncertainty — or an explicit statement that the model provides none. Creator records those, because a prediction without them is not a hypothesis, it is a guess.
What E1 does not permit: satisfying a constraint, ranking a candidate against measured alternatives, or entering design documentation as a property value. Those are rules, not guidance.
The case worth reading before you use this
In January 2025 Microsoft announced that MatterGen had designed a material to a 200 GPa bulk modulus target, synthesised at 169 GPa. In April 2026 a peer-reviewed analysis showed the compound was isostructural with one reported in 1971, present in ICSD, in the COD and in the model’s own reference dataset — and that the material actually synthesised had a different composition from the one predicted.
The full account, with sources.
That is not an argument against generative models. It is an argument for one specific check: before treating a generated composition as novel, look it up. It takes minutes in two free databases, and it was skipped precisely because the result looked good.
What Creator is for
Widening a search. Suggesting directions in composition space you had not considered, which you then evaluate by other means.
Teaching and exploration. Seeing what a model proposes for a target, and why that is not the same as a discovery.
Nothing that reaches a drawing. No output of this module belongs in design documentation without validation that is not generated by it.
What this page does not cover
- Model architecture or training data. Not documented publicly yet.
- Any accuracy claim. We have not benchmarked Creator and will not claim a number we have not measured. When we do, the benchmark will include the cases it loses.
- Synthesis routes. Creator proposes compositions, not ways to make them, and the gap between those two is where most of the difficulty lives.
- Validation. DFT or experimental confirmation is outside this tool. How far to trust computed validation.