Goldilocks ML
Setting up a DFT calculation means choosing things that are hard to choose well. How dense does the k-point mesh need to be? Is this material a metal, so it needs smearing? Too coarse and the answer is wrong; too fine and you burn compute for nothing.
Goldilocks answers those questions with models trained on past calculations.
Just want the answers?
Goldilocks Core takes a structure, fetches the right model, and writes your input files. You never touch this repository.
This site is for training and publishing the models themselves.
What you can do here
Use a model — load a published model and get a prediction, in five lines.
Train a model — describe a training job in one small configuration file and run it. Every run leaves one folder holding the predictions, the split, the scores against a baseline, and a checksum for every file it touched.
Publish a model — put it in PSDI Data Collections with a permanent identifier, so others can cite it and check they have the same file.
Models published this way
| Model | What it gives you | Record |
|---|---|---|
| QRF95 | how dense a k-point mesh needs to be | q3bye-wep37 |
| k-index forest | which mesh on the ladder a crystal needs | 4050a-aas85 |
| Metallicity classifier | metal or insulator | ba06w-n6a68 |
| CGCNN representation | 64 numbers describing a crystal | m742g-g0k14 |
All models, including the ones trained here and not yet published.