Transferability of atom-based neural networks

Machine-learning models in chemistry—when based on descriptors of atoms embedded within molecules—face essential challenges in transferring the quality of predictions of local electronic structures and their associated properties across chemical compound space. In the present work, we make use of ad...

Full description

Saved in:
Bibliographic Details
Main Authors: Frederik Ø Kjeldal, Janus J Eriksen
Format: Article
Language:English
Published: IOP Publishing 2024-01-01
Series:Machine Learning: Science and Technology
Subjects:
Online Access:https://doi.org/10.1088/2632-2153/ad9709
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1846139527258701824
author Frederik Ø Kjeldal
Janus J Eriksen
author_facet Frederik Ø Kjeldal
Janus J Eriksen
author_sort Frederik Ø Kjeldal
collection DOAJ
description Machine-learning models in chemistry—when based on descriptors of atoms embedded within molecules—face essential challenges in transferring the quality of predictions of local electronic structures and their associated properties across chemical compound space. In the present work, we make use of adversarial validation to elucidate certain intrinsic complications related to machine inferences of unseen chemistry. On this basis, we employ invariant and equivariant neural networks—both trained either exclusively on total molecular energies or a combination of these and data from atomic partitioning schemes—to evaluate how such models scale performance-wise between datasets of fundamentally different functionality and composition. We find the inference of local electronic properties to improve significantly when training models on augmented data that appropriately expose local functional features. However, molecular datasets for training purposes must themselves be sufficiently comprehensive and rich in composition to warrant any generalizations to larger systems, and even then, transferability can still only genuinely manifest if the body of atomic energies available for training purposes exposes the uniqueness of different functional moieties within molecules. We demonstrate this point by comparing machine models trained on atomic partitioning schemes based on the spatial locality of either native atomic or molecular orbitals.
format Article
id doaj-art-fe88dd4ce639402a835b4b38cdbe6836
institution Kabale University
issn 2632-2153
language English
publishDate 2024-01-01
publisher IOP Publishing
record_format Article
series Machine Learning: Science and Technology
spelling doaj-art-fe88dd4ce639402a835b4b38cdbe68362024-12-06T10:04:29ZengIOP PublishingMachine Learning: Science and Technology2632-21532024-01-015404505910.1088/2632-2153/ad9709Transferability of atom-based neural networksFrederik Ø Kjeldal0Janus J Eriksen1https://orcid.org/0000-0001-8583-3842DTU Chemistry, Technical University of Denmark , Kemitorvet Bldg. 206, 2800 Kgs. Lyngby, DenmarkDTU Chemistry, Technical University of Denmark , Kemitorvet Bldg. 206, 2800 Kgs. Lyngby, DenmarkMachine-learning models in chemistry—when based on descriptors of atoms embedded within molecules—face essential challenges in transferring the quality of predictions of local electronic structures and their associated properties across chemical compound space. In the present work, we make use of adversarial validation to elucidate certain intrinsic complications related to machine inferences of unseen chemistry. On this basis, we employ invariant and equivariant neural networks—both trained either exclusively on total molecular energies or a combination of these and data from atomic partitioning schemes—to evaluate how such models scale performance-wise between datasets of fundamentally different functionality and composition. We find the inference of local electronic properties to improve significantly when training models on augmented data that appropriately expose local functional features. However, molecular datasets for training purposes must themselves be sufficiently comprehensive and rich in composition to warrant any generalizations to larger systems, and even then, transferability can still only genuinely manifest if the body of atomic energies available for training purposes exposes the uniqueness of different functional moieties within molecules. We demonstrate this point by comparing machine models trained on atomic partitioning schemes based on the spatial locality of either native atomic or molecular orbitals.https://doi.org/10.1088/2632-2153/ad9709neural networkselectronic-structure theoryatomic decomposition schemesatomic energies
spellingShingle Frederik Ø Kjeldal
Janus J Eriksen
Transferability of atom-based neural networks
Machine Learning: Science and Technology
neural networks
electronic-structure theory
atomic decomposition schemes
atomic energies
title Transferability of atom-based neural networks
title_full Transferability of atom-based neural networks
title_fullStr Transferability of atom-based neural networks
title_full_unstemmed Transferability of atom-based neural networks
title_short Transferability of atom-based neural networks
title_sort transferability of atom based neural networks
topic neural networks
electronic-structure theory
atomic decomposition schemes
atomic energies
url https://doi.org/10.1088/2632-2153/ad9709
work_keys_str_mv AT frederikøkjeldal transferabilityofatombasedneuralnetworks
AT janusjeriksen transferabilityofatombasedneuralnetworks