Entries for these benchmarks record the mean number number of epochs
mean_epochs in benchmark.dat, median number of epochs median_epochs, mean energy rmse as
mean_energy_rmse, median energy rmse as median_energy_rmse, mean force rmse as mean_force_rmse,
median force rmse as median_force_rmse , and the number of failed minimizations as nfailed
A mean and median F_rmse value of -1.0 implies only energies were trained for that model.
This benchmark tests the performance of optimizers starting from 50 random uniform in range [-1,1) initializations of
neural network models. The structures are from a molecular dynamics simulation(MD) at 300K, of a 38 atom platinum nanoparticle.
The atoms follow the effective medium theory (EMT) potential as implemented in the atomic simulation environment (ASE)1.
A trajetory file containing the structures is here: pt_md.traj.
The L2 norm of the loss gradients per parameter set (e.g a parameter set may be the set of all input layer
biases), must be reduced to at least 0.02 energy/distance (if only energy training), or (0.1 energy/distance)
if training forces also. The maximum number of epoch that may be made cannot exceed 10,000 epochs
(if training only energies) or 30,000 (for force training as well).
The neural network architecture here is two layers with 10 and 7 neurons respectively.
The energies fitted are not scaled by any feauture scaling algorithms (e.g like that of scipy's StandardScaler).
The activation function is tanh. All the data has been partitioned into one batch for training,
and the learning rate for all the optimizers is 0.01.
If any of these defaults are changed, it will be determinable by the filename of the benchmark entry.
Runs that exceed the maximum number of epochs or fail to converge for
any other reason are considered failed.
| Entry | Mean Epochs | Median Epochs | Mean E_rmse | Median E_rmse | Mean F_rmse | Median F_rmse | Failed |
|---|---|---|---|---|---|---|---|
|
pyamff-adam10-7-fcoeff0
Date: 12 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-adam10-7-fcoeff0.tgz |
1103.24000 | 868.50000 | 0.00444 | 0.00439 | -1.00000 | -1.00000 | 0 |
|
pyamff-adam10-7-fcoeff0_1
Date: 09 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-adam10-7-fcoeff0_1.tgz |
1708.80000 | 1597.00000 | 0.00090 | 0.00090 | 0.10531 | 0.10434 | 0 |
|
pyamff-lbfgs10-7-fcoeff0
Date: 12 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-lbfgs10-7-fcoeff0.tgz |
1791.80000 | 1275.50000 | 0.00390 | 0.00363 | -1.00000 | -1.00000 | 1 |
|
pyamff-rprop10-7-fcoeff0
Date: 12 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-rprop10-7-fcoeff0.tgz |
97.58000 | 99.00000 | 0.00391 | 0.00373 | -1.00000 | -1.00000 | 0 |
|
pyamff-rprop10-7-fcoeff0_1
Date: 09 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-rprop10-7-fcoeff0_1.tgz |
1022.86000 | 901.00000 | 0.00082 | 0.00081 | 0.09959 | 0.09919 | 0 |
|
pyamff-lbfgs10-7-fcoeff0_1
Date: 09 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-lbfgs10-7-fcoeff0_1.tgz |
1943.34000 | 1403.00000 | 0.00086 | 0.00074 | 0.10037 | 0.09041 | 0 |
| Show/Hide Additional Entries... | |||||||
This benchmark is very similar to the Pt nanoparticle benchmark, above. This tests the performance of optimizers training on 570 strucutres of ethylene extracted from a density functional theory (DFT) molecular dynamics simulation. A trajectory file containing the structures is here: C2H4_md.traj.
| Entry | Mean Epochs | Median Epochs | Mean E_rmse | Median E_rmse | Mean F_rmse | Median F_rmse | Failed |
|---|---|---|---|---|---|---|---|
|
pyamff-adam10-7-fcoeff0
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-adam10-7-fcoeff0.tgz |
4395.30000 | 4106.00000 | 0.01177 | 0.01205 | -1.00000 | -1.00000 | 0 |
|
pyamff-adam10-7-fcoeff0_1
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-adam10-7-fcoeff0_1.tgz |
13876.70000 | 9683.00000 | 0.02110 | 0.01972 | 0.15999 | 0.17328 | 10 |
|
pyamff-lbfgs10-7-fcoeff0
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-lbfgs10-7-fcoeff0.tgz |
2157.14000 | 1560.00000 | 0.00813 | 0.00703 | -1.00000 | -1.00000 | 0 |
|
pyamff-rprop10-7-fcoeff0
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-rprop10-7-fcoeff0.tgz |
184.56000 | 170.50000 | 0.00863 | 0.00824 | -1.00000 | -1.00000 | 0 |
|
pyamff-rprop10-7-fcoeff0_1
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-rprop10-7-fcoeff0_1.tgz |
12926.88000 | 4493.50000 | 0.01827 | 0.01923 | 0.16424 | 0.17673 | 18 |
|
pyamff-lbfgs10-7-fcoeff0_1
Date: 13 Dec 2022 Contributor: Henkelman Group Code: pyamff-8037eb98.tgz Input files: pyamff-lbfgs10-7-fcoeff0_1.tgz |
23314.46000 | 25349.00000 | 0.01154 | 0.00986 | 0.11970 | 0.09140 | 19 |
| Show/Hide Additional Entries... | |||||||
References.
1. Larsen et. al, J. Phys.: Condens. Matter Vol. 29 273002, 2017