Note the open electrostatics gap against the paper's reported MAEs
Browse files
README.md
CHANGED
|
@@ -29,6 +29,11 @@ checkpoints, not a retrained approximation: every atom-model tensor transfers
|
|
| 29 |
bit-exactly, and predictions match recorded TensorFlow output to ~1e-6 on
|
| 30 |
multipoles and ~1e-4 kcal/mol on interaction energies (float32 reduction order).
|
| 31 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 32 |
Two things they require, both handled by `weights="ap2_tf_paper"`:
|
| 33 |
|
| 34 |
- each `pair{i}` must be loaded with its own `atom{i}` — the checkpoints carry
|
|
|
|
| 29 |
bit-exactly, and predictions match recorded TensorFlow output to ~1e-6 on
|
| 30 |
multipoles and ~1e-4 kcal/mol on interaction energies (float32 reduction order).
|
| 31 |
|
| 32 |
+
Against the paper's *reported* MAEs, the five-member average matches exchange,
|
| 33 |
+
induction and dispersion to 2-6e-4 kcal/mol and sits +0.134 kcal/mol high on
|
| 34 |
+
electrostatics, which is still open; see
|
| 35 |
+
[the parity spec](https://github.com/Awallace3/qcmlforge/blob/main/docs/specs/apnet2-tensorflow-parity.md).
|
| 36 |
+
|
| 37 |
Two things they require, both handled by `weights="ap2_tf_paper"`:
|
| 38 |
|
| 39 |
- each `pair{i}` must be loaded with its own `atom{i}` — the checkpoints carry
|