Reasoning: The rule "all categories must have a corresponding label" for multi-tagging stands. Labels must have descriptive names. Then, universal-potential would make more sense. But, the most closely related catgories, ml-iap and rep-learn, also have abbreviatory IDs. So, stick with uip. In the context of this list, people know what it means. If not, the label's URL clarifies.
Category subtitle / Label description: Machine-learned interatomic potentials (ML-IAP) that have been trained on large, chemically and structural diverse datasets. For materials, this means e.g. datasets that include a majority of the periodic table.
Issue checklist.
[x] Create category and label
[x] Move projects into category ("major label"):
[x] In projects.yaml
M3GNet, CHGNet , GNoME
[x] In open issues
242, #300, #308, #313, #314, #332, #339
Special cases.
306 calls itself "universal potential", but isn't. It does not fit the UIP category description above wrt the original publication. There, AisNet was not pre-trained on large sets of datasets.
Additional context:
Since ca. 2022, universal potentials start to appear. Some have already been added here. Some early prominent examples include M3GNet, CHGNet, MACE-MP0 and others. A good but incomplete overview is given on the Matbench Discovery leaderboard.
Care should be taken to sharpen the definition.
Possible overlaps, future conflicts with related concepts. Foundation models, multi-modal and multi-target learning models and frameworks.
Category details:
Description.
uip,universal-potentialuipuniversal-potentialwould make more sense. But, the most closely related catgories,ml-iapandrep-learn, also have abbreviatory IDs. So, stick withuip. In the context of this list, people know what it means. If not, the label's URL clarifies.uipIssue checklist.
projects.yaml242, #300, #308, #313, #314, #332, #339
306 calls itself "universal potential", but isn't. It does not fit the UIP category description above wrt the original publication. There, AisNet was not pre-trained on large sets of datasets.
Additional context:
Since ca. 2022, universal potentials start to appear. Some have already been added here. Some early prominent examples include M3GNet, CHGNet, MACE-MP0 and others. A good but incomplete overview is given on the Matbench Discovery leaderboard.
Care should be taken to sharpen the definition.
Possible overlaps, future conflicts with related concepts. Foundation models, multi-modal and multi-target learning models and frameworks.