Exploring chemical and conformational spaces by batch mode deep active learning

Zaverkin, Viktor; Holzmüller, David; Steinwart, Ingo; Kästner, Johannes

doi:10.1039/d2dd00034b

Cited by 25 publications

(25 citation statements)

References 46 publications

(262 reference statements)

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For example, our earlier works developed ensemble-free active learning approaches for interatomic neural network (NN) potentials based on the last layer and sketched gradient features. [9][10][11] These approaches provided a learned similarity measure between data points by considering the gradient kernel of a trained NN, which corresponds to the finite-width neural tangent kernel. 12 Additionally, to avoid selecting similar structures, ref.…”

Section: Introductionmentioning

confidence: 99%

Transfer learning for chemically accurate interatomic neural network potentials

Zaverkin

Holzmüller

Bonfirraro

et al. 2023

Phys. Chem. Chem. Phys.

Self Cite

View full text Add to dashboard Cite

show abstract

Section: Introductionmentioning

confidence: 99%

Transfer learning for chemically accurate interatomic neural network potentials

Zaverkin

Holzmüller

Bonfirraro

et al. 2023

Phys. Chem. Chem. Phys.

Self Cite

View full text Add to dashboard Cite

show abstract

“…AL 32,33,34 aims to iteratively collect diverse training datasets addressing any weaknesses identified in an ML model prediction. For this, it is necessary to estimate uncertainty for a model's predictions.…”

Section: Mainmentioning

confidence: 99%

“…Another feature of AL is that it can employ physically meaningful dynamical trajectories for the sampling of configurations. In the present work, we illustrate how to keep these benefits of AL, while accelerating the rate of new data collection.AL 32,33,34 aims to iteratively collect diverse training datasets addressing any weaknesses identified in an ML model prediction. For this, it is necessary to estimate uncertainty for a model's predictions.…”

mentioning

confidence: 99%

Uncertainty Driven Dynamics for Active Learning of Interatomic Potentials

Kulichenko

Barros

Lubbers

et al. 2022

Preprint

View full text Add to dashboard Cite

Machine learning (ML) models, if trained to datasets of high-fidelity quantum simulations, produce accurate and efficient interatomic potentials. Active learning (AL) is a powerful tool to iteratively generate diverse datasets. In this approach, the ML model provides an uncertainty estimate along with its prediction for each new atomic configuration. If the uncertainty estimate passes a certain threshold, then the configuration is included in the dataset. A key challenge in this process is locating structures for which the model lacks underlying training data. Here, we develop a strategy to more rapidly discover configurations that meaningfully augment the training dataset. The approach, uncertainty driven dynamics for active learning (UDD-AL), modifies the potential energy surface used in molecular dynamics simulations to favor regions of configuration space for which there is large model uncertainty. Performance of UDD-AL is demonstrated for two challenging AL tasks: sampling the conformational space of glycine and sampling the promotion of proton transfer in acetylacetone. The method is shown to efficiently explore chemically relevant configuration space, which may be inaccessible using regular dynamical sampling at target temperature conditions.

show abstract

“…45 The actual measure of the uncertainty is the determinant of the corresponding MaxVol submatrix. For GM-NN, the uncertainty is computed by employing gradient features of a trained NN (with θ θ θ being the weights and biases), 26,27 corresponding to the finite-width neural tangent kernel. 46 In this work, the largest cluster maximum distance (LCMD) method for selecting new structures and last-layer gradient features (FEAT(LL)) to compute the similarity between structures and the model's uncertainty from it is used.…”

Section: Features Of the Two ML Modelsmentioning

confidence: 99%

“…ML-based models can also robustly estimate the uncertainty in their predicted energies and forces. [22][23][24][25][26][27] Such estimations are used to develop active learning algorithms to select the data most informative to the model. The data is obtained from the configurational and vibrational spaces sampled during, e.g., a molecular dynamics (MD) simulation using the existing model, either beforehand or in an on-the-fly fashion.…”

Section: Introductionmentioning

confidence: 99%

Exploring the limits of machine-learned potentials for chemically complex multicomponent systems

Gubaev

Zaverkin

Srinivasan

et al. 2022

Preprint

View full text Add to dashboard Cite

Chemically complex multicomponent alloys have garnered widespread interest owing to their exceptional properties coming from a sheer inexhaustible compositional space. The complexity poses severe challenges for atomistic modelling and interatomic potential development. Here, we explore the limits of two complementary state-of-the-art machine-learned potentials—the moment tensor potential (MTP) and the Gaussian moment neural network (GM-NN)—in simultaneously describing both the configurational and vibrational degrees of freedom in the prototype Ta-V-Cr-W alloy family. Both models are equally accurate with exceptional performance in comparison to classical potentials. A single potential is able to achieve root-mean-square-errors of 1.37-4.35 meV/atom and 0.023-0.057 eV/Å in 0 K energies and forces, respectively, across all subsystems of the alloy family, and 0.156-0.179 eV/Å in high-temperature molecular dynamics forces for the disordered quaternary. MTPs achieve faster convergence with the training size than the GM-NNs, whereas GM-NNs are faster in execution. Active learning is partially beneficial and has to be complemented with a conventional human-based training set generation.

show abstract

Exploring chemical and conformational spaces by batch mode deep active learning

Abstract: The development of machine-learned interatomic potentials requires generating sufficiently expressive atomistic data sets. Active learning algorithms select data points on which labels, i.e., energies and forces, are calculated for inclusion...

Cited by 25 publications

References 46 publications

Transfer learning for chemically accurate interatomic neural network potentials

Transfer learning for chemically accurate interatomic neural network potentials

Uncertainty Driven Dynamics for Active Learning of Interatomic Potentials

Exploring the limits of machine-learned potentials for chemically complex multicomponent systems

Contact Info

Product

Resources

About