KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory

Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:The Journal of chemical physics 2023-10, Vol.159 (14)
Hauptverfasser: Remme, R., Kaczun, T., Scheurer, M., Dreuw, A., Hamprecht, F. A.
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue 14
container_start_page
container_title The Journal of chemical physics
container_volume 159
creator Remme, R.
Kaczun, T.
Scheurer, M.
Dreuw, A.
Hamprecht, F. A.
description Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional from ground truth provided by the more expensive Kohn–Sham density functional theory. Such learning is confronted with two key challenges: Giving the model sufficient expressivity and spatial context while limiting the memory footprint to afford computations on a GPU and creating a sufficiently broad distribution of training data to enable iterative density optimization even when starting from a poor initial guess. In response, we introduce KineticNet, an equivariant deep neural network architecture based on point convolutions adapted to the prediction of quantities on molecular quadrature grids. Important contributions include convolution filters with sufficient spatial resolution in the vicinity of nuclear cusp, an atom-centric sparse but expressive architecture that relays information across multiple bond lengths, and a new strategy to generate varied training data by finding ground state densities in the face of perturbations by a random external potential. KineticNet achieves, for the first time, chemical accuracy of the learned functionals across input densities and geometries of tiny molecules. For two-electron systems, we additionally demonstrate OF-DFT density optimization with chemical accuracy.
doi_str_mv 10.1063/5.0158275
format Article
fullrecord <record><control><sourceid>proquest_cross</sourceid><recordid>TN_cdi_crossref_primary_10_1063_5_0158275</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2876724412</sourcerecordid><originalsourceid>FETCH-LOGICAL-c325t-be0b5bb286af232fc9f1bac7ad4383b808aca026c7f66535b1a7c279dcf632f23</originalsourceid><addsrcrecordid>eNp90D1PwzAQgGELgUQpDPwDSyyAlGI7tZ2wofIpKlhgjmz3XFJSu9jO0H9PQrrAwHTLo9Pdi9ApJRNKRH7FJ4Tygkm-h0aUFGUmRUn20YgQRrNSEHGIjmJcEUKoZNMRWj_XDlJtXiBd41uADW5ABVe7JVY4BeWihaB0A_hzgBgchOUW29aZVHunGmx9wD7oOqkmswEAL8DFOv0y6QN82B6jA6uaCCe7OUbv93dvs8ds_vrwNLuZZyZnPGUaiOZas0Ioy3JmTWmpVkaqxTQvcl2QQhlFmDDSCsFzrqmShslyYazoOMvH6HzYuwn-q4WYqnUdDTSNcuDbWLFCyryQnPf07A9d-TZ0J_8o0UWa0l5dDMoEH2MAW21CvVZhW1FS9eErXu3Cd_ZysNF0Sfr__8Hfyi2D9g</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>2876724412</pqid></control><display><type>article</type><title>KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory</title><source>American Institute of Physics (AIP) Journals</source><source>Alma/SFX Local Collection</source><creator>Remme, R. ; Kaczun, T. ; Scheurer, M. ; Dreuw, A. ; Hamprecht, F. A.</creator><creatorcontrib>Remme, R. ; Kaczun, T. ; Scheurer, M. ; Dreuw, A. ; Hamprecht, F. A.</creatorcontrib><description>Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional from ground truth provided by the more expensive Kohn–Sham density functional theory. Such learning is confronted with two key challenges: Giving the model sufficient expressivity and spatial context while limiting the memory footprint to afford computations on a GPU and creating a sufficiently broad distribution of training data to enable iterative density optimization even when starting from a poor initial guess. In response, we introduce KineticNet, an equivariant deep neural network architecture based on point convolutions adapted to the prediction of quantities on molecular quadrature grids. Important contributions include convolution filters with sufficient spatial resolution in the vicinity of nuclear cusp, an atom-centric sparse but expressive architecture that relays information across multiple bond lengths, and a new strategy to generate varied training data by finding ground state densities in the face of perturbations by a random external potential. KineticNet achieves, for the first time, chemical accuracy of the learned functionals across input densities and geometries of tiny molecules. For two-electron systems, we additionally demonstrate OF-DFT density optimization with chemical accuracy.</description><identifier>ISSN: 0021-9606</identifier><identifier>EISSN: 1089-7690</identifier><identifier>DOI: 10.1063/5.0158275</identifier><identifier>CODEN: JCPSA6</identifier><language>eng</language><publisher>Melville: American Institute of Physics</publisher><subject>Artificial neural networks ; Computer architecture ; Deep learning ; Density functional theory ; Electron density ; Energy ; Ground state ; Iterative methods ; Kinetic energy ; Machine learning ; Molecular properties ; Optimization ; Perturbation ; Physics ; Quadratures ; Spatial resolution</subject><ispartof>The Journal of chemical physics, 2023-10, Vol.159 (14)</ispartof><rights>Author(s)</rights><rights>2023 Author(s). Published under an exclusive license by AIP Publishing.</rights><lds50>peer_reviewed</lds50><woscitedreferencessubscribed>false</woscitedreferencessubscribed><citedby>FETCH-LOGICAL-c325t-be0b5bb286af232fc9f1bac7ad4383b808aca026c7f66535b1a7c279dcf632f23</citedby><cites>FETCH-LOGICAL-c325t-be0b5bb286af232fc9f1bac7ad4383b808aca026c7f66535b1a7c279dcf632f23</cites><orcidid>0000-0002-5862-5113 ; 0009-0003-5703-6565 ; 0000-0003-0592-3464 ; 0009-0004-7433-7041 ; 0000-0003-4148-5043</orcidid></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><linktohtml>$$Uhttps://pubs.aip.org/jcp/article-lookup/doi/10.1063/5.0158275$$EHTML$$P50$$Gscitation$$H</linktohtml><link.rule.ids>314,776,780,790,4498,27901,27902,76126</link.rule.ids></links><search><creatorcontrib>Remme, R.</creatorcontrib><creatorcontrib>Kaczun, T.</creatorcontrib><creatorcontrib>Scheurer, M.</creatorcontrib><creatorcontrib>Dreuw, A.</creatorcontrib><creatorcontrib>Hamprecht, F. A.</creatorcontrib><title>KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory</title><title>The Journal of chemical physics</title><description>Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional from ground truth provided by the more expensive Kohn–Sham density functional theory. Such learning is confronted with two key challenges: Giving the model sufficient expressivity and spatial context while limiting the memory footprint to afford computations on a GPU and creating a sufficiently broad distribution of training data to enable iterative density optimization even when starting from a poor initial guess. In response, we introduce KineticNet, an equivariant deep neural network architecture based on point convolutions adapted to the prediction of quantities on molecular quadrature grids. Important contributions include convolution filters with sufficient spatial resolution in the vicinity of nuclear cusp, an atom-centric sparse but expressive architecture that relays information across multiple bond lengths, and a new strategy to generate varied training data by finding ground state densities in the face of perturbations by a random external potential. KineticNet achieves, for the first time, chemical accuracy of the learned functionals across input densities and geometries of tiny molecules. For two-electron systems, we additionally demonstrate OF-DFT density optimization with chemical accuracy.</description><subject>Artificial neural networks</subject><subject>Computer architecture</subject><subject>Deep learning</subject><subject>Density functional theory</subject><subject>Electron density</subject><subject>Energy</subject><subject>Ground state</subject><subject>Iterative methods</subject><subject>Kinetic energy</subject><subject>Machine learning</subject><subject>Molecular properties</subject><subject>Optimization</subject><subject>Perturbation</subject><subject>Physics</subject><subject>Quadratures</subject><subject>Spatial resolution</subject><issn>0021-9606</issn><issn>1089-7690</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2023</creationdate><recordtype>article</recordtype><recordid>eNp90D1PwzAQgGELgUQpDPwDSyyAlGI7tZ2wofIpKlhgjmz3XFJSu9jO0H9PQrrAwHTLo9Pdi9ApJRNKRH7FJ4Tygkm-h0aUFGUmRUn20YgQRrNSEHGIjmJcEUKoZNMRWj_XDlJtXiBd41uADW5ABVe7JVY4BeWihaB0A_hzgBgchOUW29aZVHunGmx9wD7oOqkmswEAL8DFOv0y6QN82B6jA6uaCCe7OUbv93dvs8ds_vrwNLuZZyZnPGUaiOZas0Ioy3JmTWmpVkaqxTQvcl2QQhlFmDDSCsFzrqmShslyYazoOMvH6HzYuwn-q4WYqnUdDTSNcuDbWLFCyryQnPf07A9d-TZ0J_8o0UWa0l5dDMoEH2MAW21CvVZhW1FS9eErXu3Cd_ZysNF0Sfr__8Hfyi2D9g</recordid><startdate>20231014</startdate><enddate>20231014</enddate><creator>Remme, R.</creator><creator>Kaczun, T.</creator><creator>Scheurer, M.</creator><creator>Dreuw, A.</creator><creator>Hamprecht, F. A.</creator><general>American Institute of Physics</general><scope>AAYXX</scope><scope>CITATION</scope><scope>8FD</scope><scope>H8D</scope><scope>L7M</scope><scope>7X8</scope><orcidid>https://orcid.org/0000-0002-5862-5113</orcidid><orcidid>https://orcid.org/0009-0003-5703-6565</orcidid><orcidid>https://orcid.org/0000-0003-0592-3464</orcidid><orcidid>https://orcid.org/0009-0004-7433-7041</orcidid><orcidid>https://orcid.org/0000-0003-4148-5043</orcidid></search><sort><creationdate>20231014</creationdate><title>KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory</title><author>Remme, R. ; Kaczun, T. ; Scheurer, M. ; Dreuw, A. ; Hamprecht, F. A.</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-c325t-be0b5bb286af232fc9f1bac7ad4383b808aca026c7f66535b1a7c279dcf632f23</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2023</creationdate><topic>Artificial neural networks</topic><topic>Computer architecture</topic><topic>Deep learning</topic><topic>Density functional theory</topic><topic>Electron density</topic><topic>Energy</topic><topic>Ground state</topic><topic>Iterative methods</topic><topic>Kinetic energy</topic><topic>Machine learning</topic><topic>Molecular properties</topic><topic>Optimization</topic><topic>Perturbation</topic><topic>Physics</topic><topic>Quadratures</topic><topic>Spatial resolution</topic><toplevel>peer_reviewed</toplevel><toplevel>online_resources</toplevel><creatorcontrib>Remme, R.</creatorcontrib><creatorcontrib>Kaczun, T.</creatorcontrib><creatorcontrib>Scheurer, M.</creatorcontrib><creatorcontrib>Dreuw, A.</creatorcontrib><creatorcontrib>Hamprecht, F. A.</creatorcontrib><collection>CrossRef</collection><collection>Technology Research Database</collection><collection>Aerospace Database</collection><collection>Advanced Technologies Database with Aerospace</collection><collection>MEDLINE - Academic</collection><jtitle>The Journal of chemical physics</jtitle></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Remme, R.</au><au>Kaczun, T.</au><au>Scheurer, M.</au><au>Dreuw, A.</au><au>Hamprecht, F. A.</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory</atitle><jtitle>The Journal of chemical physics</jtitle><date>2023-10-14</date><risdate>2023</risdate><volume>159</volume><issue>14</issue><issn>0021-9606</issn><eissn>1089-7690</eissn><coden>JCPSA6</coden><abstract>Orbital-free density functional theory (OF-DFT) holds promise to compute ground state molecular properties at minimal cost. However, it has been held back by our inability to compute the kinetic energy as a functional of electron density alone. Here, we set out to learn the kinetic energy functional from ground truth provided by the more expensive Kohn–Sham density functional theory. Such learning is confronted with two key challenges: Giving the model sufficient expressivity and spatial context while limiting the memory footprint to afford computations on a GPU and creating a sufficiently broad distribution of training data to enable iterative density optimization even when starting from a poor initial guess. In response, we introduce KineticNet, an equivariant deep neural network architecture based on point convolutions adapted to the prediction of quantities on molecular quadrature grids. Important contributions include convolution filters with sufficient spatial resolution in the vicinity of nuclear cusp, an atom-centric sparse but expressive architecture that relays information across multiple bond lengths, and a new strategy to generate varied training data by finding ground state densities in the face of perturbations by a random external potential. KineticNet achieves, for the first time, chemical accuracy of the learned functionals across input densities and geometries of tiny molecules. For two-electron systems, we additionally demonstrate OF-DFT density optimization with chemical accuracy.</abstract><cop>Melville</cop><pub>American Institute of Physics</pub><doi>10.1063/5.0158275</doi><tpages>13</tpages><orcidid>https://orcid.org/0000-0002-5862-5113</orcidid><orcidid>https://orcid.org/0009-0003-5703-6565</orcidid><orcidid>https://orcid.org/0000-0003-0592-3464</orcidid><orcidid>https://orcid.org/0009-0004-7433-7041</orcidid><orcidid>https://orcid.org/0000-0003-4148-5043</orcidid></addata></record>
fulltext fulltext
identifier ISSN: 0021-9606
ispartof The Journal of chemical physics, 2023-10, Vol.159 (14)
issn 0021-9606
1089-7690
language eng
recordid cdi_crossref_primary_10_1063_5_0158275
source American Institute of Physics (AIP) Journals; Alma/SFX Local Collection
subjects Artificial neural networks
Computer architecture
Deep learning
Density functional theory
Electron density
Energy
Ground state
Iterative methods
Kinetic energy
Machine learning
Molecular properties
Optimization
Perturbation
Physics
Quadratures
Spatial resolution
title KineticNet: Deep learning a transferable kinetic energy functional for orbital-free density functional theory
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-02-11T20%3A30%3A14IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest_cross&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=KineticNet:%20Deep%20learning%20a%20transferable%20kinetic%20energy%20functional%20for%20orbital-free%20density%20functional%20theory&rft.jtitle=The%20Journal%20of%20chemical%20physics&rft.au=Remme,%20R.&rft.date=2023-10-14&rft.volume=159&rft.issue=14&rft.issn=0021-9606&rft.eissn=1089-7690&rft.coden=JCPSA6&rft_id=info:doi/10.1063/5.0158275&rft_dat=%3Cproquest_cross%3E2876724412%3C/proquest_cross%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=2876724412&rft_id=info:pmid/&rfr_iscdi=true