A Split-Architecture Machine Learning Model for Scalable HOMO–LUMO Gap Prediction: rCEG
- 1California State University, Fresno , , 2555 E. San Ramon Avenue , , ,
Abstract
The difference in energy between the highest occupied molecular orbital and lowest unoccupied molecular orbital, or the frontier orbital gap, has a variety of applications in industries ranging from drug design and molecular electronics to organic photovoltaics and photocatalysis. This energy gap governs various chemical processes including optical absorption, chemical reactivity, and electron transport. Conventional quantum chemical calculations become costly when large molecular libraries must be screened. Thus, a split-architecture machine learning ensemble for a reference-aligned HOMO–LUMO gap, the Rapidly Calculating Energy Gap (rCEG) ML model, was developed. This model makes predictions using a large collection of decision trees (ExtraTrees, RandomForest, and HistGradientBoosting), combining them at the end with ridge regression. Unlike most earlier machine learning models, which are trained on a single narrow class of molecules and tend to lose accuracy when applied outside that class, rCEG is explicitly designed to generalize across chemically diverse molecules by routing each molecule to a specialized expert submodel. The model accesses the molecular structure, along with various descriptors of the molecule. The rCEG model was trained and validated using chemical data and structures extracted from the Pitt Quantum Repository (PQR), which contains calculated HOMO and LUMO values of over 100,000 molecules and was supplemented with additional calculations using Psi4. Training used a split-architecture method where each molecule was assigned a mutually exclusive chemical domain defined by elemental composition, formal charge, radical electron count, and heavy atom count before being routed within its domain by a single descriptor threshold. The final model achieved a mean absolute error of 0.338 eV, demonstrating that model specialization can lead to fast, accurate, and scalable predictions of the frontier orbital energy gaps.
Related
- A Lightweight Universal Machine-Learning Interatomic Potential via Knowledge Distillation for Scalable Atomistic Simulations
- A Machine Learning Approach to logBB Prediction Combining Molecular Docking and Quantum-Chemical Descriptors
- Accelerating Proton Affinity Prediction with Multi-Fidelity Machine Learning
- Accelerating Phase Prediction via CALPHAD-Informed Machine Learning and Data Augmentation
- Machine-learning prediction of optical band gaps and photocatalytic performance in rare-earth-doped spinel ferrites
BibTeX
@article{wang2026splitarchitecture,
title = {{A Split-Architecture Machine Learning Model for Scalable HOMO–LUMO Gap Prediction: rCEG}},
author = {Isaac Wang and Kristina D. Closser},
journal = {ACS Omega},
year = {2026},
month = oct,
doi = {10.1021/acsomega.6c09487},
url = {https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c09487/5460974/A-Split-Architecture-Machine-Learning-Model-for},
note = {CC BY-NC-ND 4.0},
}
RIS
TY - JOUR
TI - A Split-Architecture Machine Learning Model for Scalable HOMO–LUMO Gap Prediction: rCEG
AU - Isaac Wang
AU - Kristina D. Closser
PY - 2026
DA - 2026/10/29
JO - ACS Omega
DO - 10.1021/acsomega.6c09487
UR - https://pubs.acs.org/acsodf/article/doi/10.1021/acsomega.6c09487/5460974/A-Split-Architecture-Machine-Learning-Model-for
AB - The difference in energy between the highest occupied molecular orbital and lowest unoccupied molecular orbital, or the frontier orbital gap, has a variety of applications in industries ranging from drug design and molecular electronics to organic photovoltaics and photocatalysis. This energy gap governs various chemical processes including optical absorption, chemical reactivity, and electron transport. Conventional quantum chemical calculations become costly when large molecular libraries must be screened. Thus, a split-architecture machine learning ensemble for a reference-aligned HOMO–LUMO gap, the Rapidly Calculating Energy Gap (rCEG) ML model, was developed. This model makes predictions using a large collection of decision trees (ExtraTrees, RandomForest, and HistGradientBoosting), combining them at the end with ridge regression. Unlike most earlier machine learning models, which are trained on a single narrow class of molecules and tend to lose accuracy when applied outside that class, rCEG is explicitly designed to generalize across chemically diverse molecules by routing each molecule to a specialized expert submodel. The model accesses the molecular structure, along with various descriptors of the molecule. The rCEG model was trained and validated using chemical data and structures extracted from the Pitt Quantum Repository (PQR), which contains calculated HOMO and LUMO values of over 100,000 molecules and was supplemented with additional calculations using Psi4. Training used a split-architecture method where each molecule was assigned a mutually exclusive chemical domain defined by elemental composition, formal charge, radical electron count, and heavy atom count before being routed within its domain by a single descriptor threshold. The final model achieved a mean absolute error of 0.338 eV, demonstrating that model specialization can lead to fast, accurate, and scalable predictions of the frontier orbital energy gaps.
SN - 2470-1343
ER -