QM9 results. Results for TSNet and TS-shared when trained upon the QM9. Loss values are coordinate angstroms (Å) unless otherwise stated, while distance error values are Euclidean distance Å. Both TSNet and TSNet-shared perform exceptionally well on the midpoint prediction task, producing validation prediction indistinguishable from “true” target structures. TSNet performance predicting internal energy values is over an order of magnitude worse than state-of-the-art networks. Effort was not put into constructing a model capable of producing high accuracy energy predictions, as the ultimate goal was for transition state geometry prediction.
| Model | Train | Validation | ||
|---|---|---|---|---|
| Mean loss | Mean distance error | Mean loss | Mean distance error | |
| TSNet (midpoint) | 0.1286 | 0.03893 | 0.1355 | 0.05694 |
| TSNet-shared (midpoint) | 0.1962 | 0.03589 | 0.2209 | 0.06257 |
| TSNet (energy) | 0.1575 (eV) | — | 0.3976 (eV) | — |