What Does a Semilocal Machine-Learning Correction Actually Learn? Size-Dependent Errors across Four Parent…

Abstract

Machine-learning (ML) corrections to density functional approximations (DFAs) offer a route to improving electronic-structure predictions while retaining the efficiency of semilocal functionals. An important question is whether such corrections learn transferable improvements or instead compensate for errors specific to the parent functional. Here, we address this question by applying the semilocal ML correction of Wang \textit{et al.} [J.~Chem.~Phys.~\textbf{158}, 154107 (2023)] to PBE, B3LYP, SCAN, and r²SCAN, while keeping the network architecture, loss function, training set, and optimization protocol unchanged. We find that the learned correction develops a systematic size-dependent contribution to atomization energies for all four parents, with |r|≥0.9999 along the n-alkane series, where r is the Pearson coefficient of the linear fit. Its magnitude and sign depend systematically on the parent: the correction partially compensates the size-dependent errors of PBE and B3LYP, but introduces substantial size-dependent contributions for SCAN and r²SCAN, whose parent errors show little size dependence. In contrast, ionization potentials, electron affinities, and isomerization energies are largely unaffected. Spatial decomposition further reveals distinct origins of this extensive contribution, with bonding regions dominating for B3LYP, SCAN, and r²SCAN and core regions dominating for PBE. These results show that the global-energy ML correction is strongly dependent on the error structure of its parent DFA, highlighting size-dependent error accumulation as a critical consideration in the assessment and development of ML-enhanced density functionals.

Where machine learning for enantioselectivity on polysaccharide chiral stationary phases actually fails: a leave-one-paper-out reappraisal of 10,525 enantiomer pairs
Lang Cui
ChemRxiv4 Nov 2026PreprintLicence not statedGeneral chemistryChemRxiv
Dysfunctional Functionals and their Therapies in Time-Dependent Density Functional Theory
Neepa T. Maitra
arXiv1 Nov 2026PreprintCC BY 4.0Full textPhysical chemistry Condensed matterarXiv
Pushing the accuracy of on-top functionals with agent-driven supervised learning
Yuhao Chen, Dayou Zhang, Donald G. Truhlar +1
arXiv28 Sep 2026PreprintarXiv licenceFull textPhysical chemistryarXiv
RS-CIDER: A non-local machine learning model for approximating screened hybrid functionals
Zhuotao Jin, Mohamed S. Abdallah, Boris Kozinsky +1
arXiv11 Sep 2026PreprintarXiv licenceFull textPhysical chemistry Condensed matterarXiv
The Thiolate Shell as a Dynamic Gold Reservoir: Size-Dependent Thermal Restructuring of Gold Nanoclusters from MACE-Based Machine-Learning Molecular Dynamics
Hoang-Duc Le-Vu, Cuong-Ha Minh, Dominik Domin +2
ChemRxiv5 Nov 2026PreprintLicence not statedGeneral chemistryChemRxiv

Cite

BibTeX RIS

BibTeX

@misc{bhattacharjee2026what,
  title = {{What Does a Semilocal Machine-Learning Correction Actually Learn? Size-Dependent Errors across Four Parent Functionals}},
  author = {Abhishek Bhattacharjee and Kishan Kumar Mohanta and Subrata Jana and Prasanjit Samal},
  howpublished = {arXiv},
  year = {2026},
  month = sep,
  doi = {10.48550/arxiv.2609.37571},
  eprint = {2609.37571},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.37571},
}

RIS

TY  - UNPB
TI  - What Does a Semilocal Machine-Learning Correction Actually Learn? Size-Dependent Errors across Four Parent Functionals
AU  - Abhishek Bhattacharjee
AU  - Kishan Kumar Mohanta
AU  - Subrata Jana
AU  - Prasanjit Samal
PY  - 2026
DA  - 2026/09/29
PB  - arXiv
DO  - 10.48550/arxiv.2609.37571
UR  - https://arxiv.org/abs/2609.37571
AB  - Machine-learning (ML) corrections to density functional approximations (DFAs) offer a route to improving electronic-structure predictions while retaining the efficiency of semilocal functionals. An important question is whether such corrections learn transferable improvements or instead compensate for errors specific to the parent functional. Here, we address this question by applying the semilocal ML correction of Wang \textit{et al.} [J.~Chem.~Phys.~\textbf{158}, 154107 (2023)] to PBE, B3LYP, SCAN, and r^2SCAN, while keeping the network architecture, loss function, training set, and optimization protocol unchanged. We find that the learned correction develops a systematic size-dependent contribution to atomization energies for all four parents, with |r|≥0.9999 along the n-alkane series, where r is the Pearson coefficient of the linear fit. Its magnitude and sign depend systematically on the parent: the correction partially compensates the size-dependent errors of PBE and B3LYP, but introduces substantial size-dependent contributions for SCAN and r^2SCAN, whose parent errors show little size dependence. In contrast, ionization potentials, electron affinities, and isomerization energies are largely unaffected. Spatial decomposition further reveals distinct origins of this extensive contribution, with bonding regions dominating for B3LYP, SCAN, and r^2SCAN and core regions dominating for PBE. These results show that the global-energy ML correction is strongly dependent on the error structure of its parent DFA, highlighting size-dependent error accumulation as a critical consideration in the assessment and development of ML-enhanced density functionals.
ER  -

Abstract

Related

Cite