Skip to content
DraconiticResearch
Search Sources API
203,203 studies31 sourcesUpdated 8 min ago
PreprintCC BYChemRxiv

The Illusion of Chemical Continuity Limits Machine Learning Generalization

  • Zhibo Wang1,2
  • Alexander Davis1,2
  • Rajarshi Dutta3
  • Alex-Cristian Tomut2
  • Kanishk Yadav4
  • Ihor Neporozhnii1,2
  • Sjoerd Hoogland1,2
  • Oleksandr Voznyy1,2
4 affiliations
  1. 1Alliance for AI-Accelerated Materials Discovery
  2. 2University of Toronto
  3. 3Indian Institute of Technology Kanpur
  4. 4Birla Institute of Technology and Science, Pilani
Open at sourceDOICite

Abstract

The central dogma of materials informatics is the "Scaling Hypothesis": the assumption that chemical space is a continuous manifold, and that, with sufficient data, deep learning models can approximate any structure-property relationship. While this strategy has achieved near-DFT accuracy for thermodynamic properties like formation energy, prediction accuracy for bandgaps suffers from a fundamental, non-converging error floor (~0.4 eV). By analyzing a comprehensive dataset of binary semiconductors and state-of-the-art Graph Neural Networks (including coGN and foundational embeddings), this work isolates the root cause: unlike the smooth, transferable landscape of total energy, the electronic bandgap manifold is topologically disjoint. Consequently, increasing dataset size fails to improve predictions, instead resulting in negative learning, where increasing the structural diversity of the training set (e.g., adding rocksalt data to zincblende models) degrades prediction accuracy for unseen chemistries. This can be traced back to the discrete quantization of atomic orbitals and the resulting band crossings, which create non-differentiable cusps in the property surface. This work establishes a boundary condition for the "Big Data" paradigm: simple regression on atomic coordinates cannot resolve the intrinsic discontinuities of eigenvalue phenomena. Instead, a more detailed integration of quantum mechanics is required, such as direct entrainment of Hamiltonians as descriptors.

Related

  1. Reliable Structure-Processing-Performance Learning in Organic Photovoltaics: Chemical Generalization, Processing-Aware Prediction and Uncertainty under Domain Shift
    DENNIS OBINNA ORJI
    ChemRxivPreprintCC BYFull textMaterials & polymersChemRxiv
  2. Polymer Informatics Atlas: Property-Specific Representations, Chemistry-Aware Generalization, and Reliability-Aware Machine Learning Across Five Polymer Properties
    Dennis Obinna Orji
    ChemRxivPreprintCC BYFull textMaterials & polymersChemRxiv
  3. A Machine Learning Approach to logBB Prediction Combining Molecular Docking and Quantum-Chemical Descriptors
    Lilian Rouille
    ChemRxivPreprintLicence not statedGeneral chemistryChemRxiv
  4. Artificial Intelligence- and Machine Learning-Driven Strategies for Catalyst Design and Sustainable Chemical Processes
    Amra Bratovčić, Vesna Tomašić
    ProcessesPeer-reviewedCC BY10 citedMaterials & polymersOpenAlex
  5. Applicability-Domain-Aware Evaluation of Machine-Learning NMR Chemical-Shift Prediction in Small drug Molecules: GT-NMR for 13C and CASCADE for 1H
    Diba Allameh Zadeh
    ChemRxivPreprintCC BYFull textGeneral chemistryChemRxiv

Cite

BibTeXRIS
BibTeX
@misc{wang2026illusion,
  title = {{The Illusion of Chemical Continuity Limits Machine Learning Generalization}},
  author = {Zhibo Wang and Alexander Davis and Rajarshi Dutta and Alex-Cristian Tomut and Kanishk Yadav and Ihor Neporozhnii and Sjoerd Hoogland and Oleksandr Voznyy},
  howpublished = {ChemRxiv},
  year = {2026},
  month = oct,
  doi = {10.26434/chemrxiv-2025-vj4fk/v2},
  url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-vj4fk/v2},
  note = {CC BY},
}
RIS
TY  - UNPB
TI  - The Illusion of Chemical Continuity Limits Machine Learning Generalization
AU  - Zhibo Wang
AU  - Alexander Davis
AU  - Rajarshi Dutta
AU  - Alex-Cristian Tomut
AU  - Kanishk Yadav
AU  - Ihor Neporozhnii
AU  - Sjoerd Hoogland
AU  - Oleksandr Voznyy
PY  - 2026
DA  - 2026/10/21
PB  - ChemRxiv
DO  - 10.26434/chemrxiv-2025-vj4fk/v2
UR  - https://chemrxiv.org/doi/full/10.26434/chemrxiv-2025-vj4fk/v2
AB  - The central dogma of materials informatics is the "Scaling Hypothesis": the assumption that chemical space is a continuous manifold, and that, with sufficient data, deep learning models can approximate any structure-property relationship. While this strategy has achieved near-DFT accuracy for thermodynamic properties like formation energy, prediction accuracy for bandgaps suffers from a fundamental, non-converging error floor (~0.4 eV). By analyzing a comprehensive dataset of binary semiconductors and state-of-the-art Graph Neural Networks (including coGN and foundational embeddings), this work isolates the root cause: unlike the smooth, transferable landscape of total energy, the electronic bandgap manifold is topologically disjoint. Consequently, increasing dataset size fails to improve predictions, instead resulting in negative learning, where increasing the structural diversity of the training set (e.g., adding rocksalt data to zincblende models) degrades prediction accuracy for unseen chemistries. This can be traced back to the discrete quantization of atomic orbitals and the resulting band crossings, which create non-differentiable cusps in the property surface. This work establishes a boundary condition for the "Big Data" paradigm: simple regression on atomic coordinates cannot resolve the intrinsic discontinuities of eigenvalue phenomena. Instead, a more detailed integration of quantum mechanics is required, such as direct entrainment of Hamiltonians as descriptors.
ER  -