| --- |
| library_name: pytorch |
| tags: |
| - polymers |
| - property-prediction |
| - graph-transformer |
| - polyedit |
| --- |
| |
| # PolyEdit MIPS-retrained |
|
|
| This is a leakage-controlled retraining of the public MIPS architecture for the eight |
| PolyEdit DFT properties: `Egc`, `Egb`, `Eea`, `Ei`, `EPS`, `Nc`, `Xc`, and `Eat`. |
| It is **not** a checkpoint released by the MIPS authors. The public repository does not |
| include its required MAE or downstream checkpoints, so weights were trained from scratch |
| on PolyEdit train components. |
|
|
| The model uses MIPS star-link polymer graphs, complete-graph path features, the 200-D |
| molecular-descriptor modality, the 512-D 3D atom-pair modality, and the public |
| `KfuseGraphTransformer` with cross-attention fusion. The split contains 1,300 train, |
| 75 validation, and 953 held-out polymer records with zero connected-component overlap. |
|
|
| The strict target tolerances are much narrower than this retrained evaluator's held-out |
| errors. When the true reference targets themselves are scored, only 10.996% fall inside |
| the requested tolerance. MIPS-based generation numbers must therefore be presented as |
| evaluator hit rates alongside this calibration, not as ground-truth chemical success. |
|
|
| Upstream code: <https://github.com/wjxts/MIPS> |
|
|
| PolyEdit implementation and full metrics: <https://github.com/promotion-kim/POLYEDIT/tree/tsyou/balanced-polymer-baseline-eval> |
|
|
