sGDML Symmetric Gradient Domain Machine Learning
Ab initio accelerated.
Accurate global machine learning force fields with hundreds of atoms
Learn more Get startedsubject Articles
Machine Learning of Accurate Energy-Conserving Molecular Force Fields
Chmiela, S., Tkatchenko, A., Sauceda, H. E., Poltavsky, I., Schütt, K. T., Müller, K.-R., Science Advances, 3(5), 2017, e1603015.
Towards Exact Molecular Dynamics Simulations with Machine-Learned Force Fields
Chmiela, S., Sauceda, H. E., Müller, K.-R., Tkatchenko, A., Nature Communications, 9(1), 2018, 3887.
Accurate Global Machine Learning Force Fields for Molecules with Hundreds of Atoms
Chmiela, S., Vassilev-Galindo, V., Unke, O. T., Kabylda, A., Sauceda, H. E., Tkatchenko, A., Müller, K.-R., Science Advances, 9(2), 2023, eadf0873.
sGDML: Constructing Accurate and Data Efficient Molecular Force Fields Using Machine Learning
Chmiela, S., Sauceda, H. E., Poltavsky, I., Müller, K.-R., Tkatchenko, A., Computer Physics Communications, 240, 2019, pp. 38-45.
Dynamical Strengthening of Covalent and Non-Covalent Molecular Interactions by Nuclear Quantum Effects at Finite Temperature
Sauceda, H. E., Vassilev-Galindo, V., Chmiela, S., Müller, K.-R., Tkatchenko, A., Nature Communications, 12(1), 2021, 442.
Molecular Force Fields with Gradient-domain Machine Learning (GDML): Comparison and Synergies with Classical Force Fields
Sauceda, H. E., Gastegger, M., Chmiela, S., Müller, K.-R., Tkatchenko, A., The Journal of Chemical Physics, 153, 2020, 124109.
Molecular Force Fields with Gradient-Domain Machine Learning: Construction and Application to Dynamics of Small Molecules with Coupled Cluster Forces
Sauceda, H. E., Chmiela, S., Poltavsky, I., Müller, K.-R., Tkatchenko, A., The Journal of Chemical Physics, 150, 2019, 114102.
Construction of Machine Learned Force Fields with Quantum Chemical Accuracy: Applications and Chemical Insights
Sauceda, H. E., Chmiela, S., Poltavsky, I., Müller, K.-R., Tkatchenko, A., In: Machine Learning Meets Quantum Physics, Lecture Notes in Physics (Springer), 968, 2020, pp. 277-307.
Ensemble Learning of Coarse-Grained Molecular Dynamics Force Fields with a Kernel Approach
Wang, J., Chmiela, S., Müller, K.-R., Noé, F., Clementi, C., The Journal of Chemical Physics, 152, 2020, 194106.
BIGDML - Towards Accurate Quantum Machine Learning Force Fields for Materials
Sauceda, H. E., Gálvez-González, L. E., Chmiela, S., Paz-Borbó, L. O. , Müller, K.-R., Tkatchenko, A., Nature Communications, 9(1), 2018, 3887.
code Code (latest: v1.0.3)
Replicate our numerical results or reconstruct a force field from your own dataset with a Python implementation of sGDML.
Get started
The sgdml package uses a proprietary dataset format, but it is easy to convert from and to Extended XYZ files and other popular file formats (learn more):
$ sgdml_dataset_via_ase.py <xyz_dataset_file>A force field is created via a single command-line call that yields a ready-to-use model file:
$ sgdml all <sgdml_dataset_file> <n_train> <n_validate> [<n_test>]The last three parameters specify the sizes for the training, validation and test dataset splits, which are sampled from the provided dataset file without overlap. Leave out <n_test> to use all remaining points for testing (learn more).
A force field model is effectively a parametrization of your dataset that provides energy e and forces f for any input geometry r (learn more):
import numpy as np
from sgdml.predict import GDMLPredict
from sgdml.utils import io
model = np.load('model.npz')
gdml = GDMLPredict(model)
r,_ = io.read_xyz('geometry.xyz')
e,f = gdml.predict(r)This flexibility enables many applications, e.g. by interfacing to Atomic Simulation Environment (ASE). Here are a few examples:
Training service Experimental
We offer an experimental model training service for anyone without sufficient compute resources. Simply upload your dataset, schedule some training jobs and return later to collect your model files:
file_download Datasets
| Name | Size | Benchmark | Download |
|
DFT [FHI-aims, light tier 1]
|
|||
| Benzene (Chmiela et al., 2018) | 49,863 | ||
|
MD17 dataset (Chmiela et al., 2017)
|
|||
| Benzene | 627,983 | ||
| Uracil | 133,770 | ||
| Naphthalene | 326,250 | ||
| Aspirin | 211,762 | ||
| Salicylic acid | 320,231 | ||
| Malonaldehyde | 993,237 | ||
| Ethanol | 555,092 | ||
| Toluene | 442,790 | ||
| Paracetamol | 106,490 | ||
| Azobenzene | 99,999 | ||
|
MD22 dataset (Chmiela et al., 2023)
|
|||
|
DFT [FHI-aims, tight tiers 1&2]
|
|||
| Ac-Ala3-NHMe | 85,109 | ||
| Docosahexaenoic acidDHA | 69,753 | ||
| Stachyose | 27,272 | ||
| DNA base pair (AT-AT) | 20,001 | ||
| DNA base pair (AT-AT-CG-CG) | 10,153 | ||
|
DFT [FHI-aims, light tier 1]
|
|||
| Buckyball catcher | 6,102 | ||
| Double-walled nanotubeDWNT | 5,032 | ||
CCSD [Psi4, cc-pVDZ] |
|||
| Aspirin | 1,500 | ||
CCSD(T) [Psi4, cc-pVDZ] |
|||
| Benzene | 1,500 | ||
| Malonaldehyde | 1,500 | ||
| Toluene | 1,500 | ||
CCSD(T) [Psi4, cc-pVTZ] |
|||
| Ethanol | 2,000 | ||