Automated Bandgap Correction in Double Perovskites: An Active Learning Framework Using Hybrid Functional Density Functional Theory and Delta Machine Learning

One of the major challenges that continue to exist in the area of computational materials discovery is accurate determination of the band gap, particularly when large candidate spaces must be screened at a level of fidelity suitable for decision-making. This issue is relevant to oxide double-perovskites A2BB′O6, due to their significant chemical and structural diversity, making them very attractive for energy applications; however, their evaluation is also extremely difficult using low-cost electronic structure methods alone. Semilocal Density Functional Theory (DFT) methods, such as PBE, are computationally efficient and thus useful for performing large scale screening. However, they systematically underestimate band gaps and become less reliable as the metal-nonmetal boundary is approached. Hybrid functional methods, such as HSE06, significantly improve band gap predictions over DFT, but their computational expense means that using them exhaustively to evaluate a broad range of chemical space is not practical. This dissertation presents a hybrid-level band gap prediction methodology for oxide double-perovskites, that utilizes a combination of ∆-Machine Learning and Active Learning to achieve HSE06-level band gap prediction within a limited HSE06 budget. The central strategy combines ∆-Machine Learning with Active Learning so that expensive hybrid-functional calculations are used selectively rather than uniformly. Instead of training directly on dense HSE06 coverage of a full candidate space, the methodology uses a broadly available semilocal baseline to learn the correction to the hybrid-functional target and then iteratively refines that correction via targeted acquisition of informative new HSE06 labels. A unique aspect of this research is the construction and utilization of a unified oxide double-perovskite candidate universe that does not rely upon chemically convenient preselection. The explored space was built by combining reported Materials Project A2BB′O6 structures with Glazer prototype–generated compounds, thereby extending the search beyond previously reported entries while preserving a consistent structural framework for screening. Equally important, the workflow does not remove or reduce the difficult near-zero or integer-zero PBE band gap regime from the candidate space. Rather than restricting the study to clear nonmetals, the dissertation explicitly retains these challenging systems and incorporates a regime-aware treatment for cases in which straightforward correction learning becomes unreliable. This makes the resulting predictor more realistic for discovery-scale use, where ambiguous or borderline systems cannot simply be ignored. Another contribution of this dissertation is the development and public release of DeltAL_Material, a reusable software workflow for hybrid-functional band gap correction on Materials Project-derived candidate universes within a limited labeling budget. Although the scientific validation presented in this thesis is based on oxide A2BB′O6 double perovskites, the released workflow is constructed to enable adaptation to other candidate families with the necessary feature construction, baseline definitions, and domain specific validation. The proposed active-learning policy consistently outperformed a matched random-selection baseline in terms of label efficiency. The stabilized end-to-end validation mean absolute error (MAE) decreased from 0.262 eV to 0.232 eV, and the nonmetal ∆-validation MAE decreased from 0.153 eV to 0.098 eV. In the regime in which correction learning is most effective, the average pooled nonmetal validation error remained at approximately 0.10 eV. When evaluated on a completely untouched test set, the selected model produced MAE = 0.132 eV, RMSE = 0.215 eV, and R2 = 0.977 for n = 30 materials. Therefore, the dissertation demonstrates two linked contributions. First, it demonstrates that HSE06-level band gap prediction for oxide double-perovskites can be performed in a more efficient manner by utilizing ∆-learning with selective Active Learning under a limited hybrid-functional budget. Second, it delivers DeltAL_Material, a publicly released and reproducible software workflow for HSE-level band gap correction on Materials Project-derived candidate sets.

Name thesis:degree_name

Doctor of Philosophy (PhD)

Discipline thesis:degree_discipline

Chemistry

Grantor

Science

Year dc:date.issued

2026

Author dc:creator

• Shahmohamadi, Hatef

Advisors dc:contributor.advisor

• Salahub, Dennis
• Shi, Yujun

Committee members dc:contributor.committeemember

• Kusalik, Peter
• Kennepohl, Pierre
• Achari, Gopal
• Tse, John

• Density Functional Theory
• Oxide Double Perovskites
• HSE06
• Delta Machine Learning
• Active Learning

Statement dc:rights

• Unless otherwise indicated, this material is protected by copyright and has been made available with authorization from the copyright owner. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.

Language dc:language.iso

en

DOI

https://dx.doi.org/10.11575/PRISM/51439

OAI identifier oai:identifier

oai:ucalgary.scholaris.ca:1880/124899