MACE-ANI-CC is a neural network potential (NNP) method from Dávid Péter Kovács, Ilyes Batatia, and co-workers in Gábor Csányi's lab at Cambridge in 2023.
MACE-ANI-CC is released under the MIT license. Read the MACE-ANI-CC paper. See the MACE-ANI-CC code.
| About MACE-ANI-CC | |
|---|---|
| Architecture | MACE |
| Dataset | ANI-1ccx |
| Dataset Level of Theory | CCSD(T)*/CBS |
| Dataset Size | 500k |
GMTKN55 (download) is a high-accuracy molecular benchmark that measures basic properties, reaction energies, and noncovalent interactions. The GMTKN55 benchmark comprises 55 subsets. Each subset contains a number of relative energy measurements computed at the CCSD(T) level of theory. When viewing subset-level results, the mean absolute deviation/error (MAD/MAE) in kcal/mol is shown. Each subset is assigned a difficulty weight (see Section 4). When viewing category-level results, the weighted total mean absolute deviation (using the WTMAD-2 weights) is shown. See all GMTKN55 results.
| Name | Incomplete Subsets | Benchmarked By | ||||||
|---|---|---|---|---|---|---|---|---|
| ωB97M-D3BJ/def2-QZVP | 2.86 | 5.77 | 2.34 | 4.54 | 3.63 | 4.04 | link | |
| B3LYP-D3BJ/def2-QZVP | 7.99 | 10.16 | 4.11 | 5.65 | 4.82 | 6.18 | link | |
| MACE-ANI-CC | 7.67 | 13.78 | 3.02 | 9.28 | 13 | Ari 2024-12-23 | ||
| B97-3c | 11.99 | 10.51 | 9.17 | 8.62 | 11.89 | 10.16 | link | |
| GFN2-xTB | 20.02 | 19.77 | 13.66 | 24.58 | 11.44 | 18.65 | link |
The Folmsbee conformers benchmark (GitHub) is a high-accuracy molecular benchmark that evaluates the accuracy of predicting and ranking conformer relative energies. The benchmark contains 708 subsets of 10 conformers each, 632 of which have energies computed at the DLPNO-CCSD(T) level of theory. DLPNO-CCSD(T) calculations were not finished for 76 subsets, so every level of theory will show at least 76 incomplete subsets. Mean absolute error (MAE) and root mean square error (RMSE) are shown in kcal/mol. See all Folmsbee results.
| Name | Incomplete Subsets | Benchmarked By | ||||
|---|---|---|---|---|---|---|
| ωB97X-D/def2-TZVP | 0.24 | 0.30 | 0.83 | 0.85 | 80 | link |
| B97-3c | 0.30 | 0.37 | 0.80 | 0.82 | 77 | link |
| MACE-ANI-CC | 0.32 | 0.40 | 0.77 | 0.79 | 95 | Ari 2025-01-15 |
| GFN2-xTB | 0.72 | 0.88 | 0.57 | 0.59 | 76 | link |
TorsionNet206 is a high-accuracy molecular benchmark that evaluates the accuracy of predicting and ranking of dihedral energy profiles. The benchmark contains 206 subsets of 24 conformers each computed at the CCSD(T)/def2-TZVP level of theory. Mean absolute error (MAE) and root mean square error (RMSE) are shown in kcal/mol. See all TorsionNet206 results.
| Name | Incomplete Subsets | Benchmarked By | ||||
|---|---|---|---|---|---|---|
| ωB97M-D3BJ/vDZP | 0.15 | 0.18 | 0.99 | 0.98 | C.W. | |
| B97-3c | 0.35 | 0.45 | 0.98 | 0.97 | C.W. | |
| MACE-ANI-CC | 0.23 | 0.29 | 0.97 | 0.97 | 59 | Ari 2025-01-11 |
| GFN2-xTB | 0.73 | 0.91 | 0.85 | 0.85 | C.W. |
This is a benchmark that evaluates the average speed of running energy and force calculations on 10-, 100-, and 1000-atom molecular systems. All calculations were run on Nvidia A10G GPUs through Modal. Click items in legend to show/hide. See all molecular speed results.
This is a benchmark that evaluates the average speed of running energy and force calculations on 10-, 100-, and 1000-atom periodic systems. All calculations were run on Nvidia A10G GPUs through Modal. Click items in legend to show/hide. See all periodic speed results.