At-Scale Discovery of High-Voltage Cathode Active Materials for Na-Ion Batteries via Charged-Only Structures and Committee Divide-and-Conquer Learning
Abstract
Current cathode-active materials (CAMs) designed for sodium-ion batteries (SIBs) suffer from low operating voltages, limited energy density, irreversible phase transitions, and low intrinsic electrical conductivity. In this work, we have developed generalizable and transferable descriptor-based machine learning (ML) models, along with composition- and graph-based neural networks (NNs) and transfer learning (TL) models, all trained on a dataset of charged-only structures rather than charged/discharged pairs. Using a committee of the top four trained ML models, the mean voltage is predicted as the target property. More importantly, we show that uncertainty quantification can improve the reliability of candidate selection for training subsequent higher-generation models by including structures for which the current committee models are least confident. In this way, the new expensive ab initio data provide information in regions where the original models most need improvement. We have showcased the transferability of the final generation models by applying them to a large database of Na-free inorganic host structures curated from Materials Project, AFLOW, OQMD, and GNoME for at-scale SIB CAM discovery. A comprehensive interactive online database, NaCAM-Atlas, was also developed, containing 14,843 CAMs with voltages greater than 3 V for SIBs. Finally, a subset of the top-ranked candidate CAMs was validated through explicit, high-throughput first-principles calculations of voltage profiles, structural robustness upon sodiation/desodiation, and electronic properties. Together, this integrated data curation, ML training and improvement, and first-principles validation strategy establishes a scalable and transferable framework for accelerating the discovery of stable, high-voltage CAMs for SIBs and beyond.
Related
- Learning the Chronology of Materials Discovery for Synthesizability Prediction
- SCALE: Simulation-Calibrated Amortized Learning for Energy Materials (A hybrid architecture connecting deterministic modeling, real-world data, and transformer-scale inference for accelerated energy-materials discovery)
- Organizational Learning in Materials Discovery Campaigns: Which Databases, Calculations, and Agent Teams Pay Off
- Synergistic effect of Mn substitution and ball milling on NaCu 0.2 Fe 0.8− x Mn x O 2 cathode materials for sodium-ion batteries
- Reaction-induced Localized Annealing for On-chip Materials Discovery
BibTeX
@misc{akter2026atscale,
title = {{At-Scale Discovery of High-Voltage Cathode Active Materials for Na-Ion Batteries via Charged-Only Structures and Committee Divide-and-Conquer Learning}},
author = {Suchona Akter and Md Omar Faruque and Mohammad R. Momeni},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2605.27229},
eprint = {2605.27229},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.27229},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - At-Scale Discovery of High-Voltage Cathode Active Materials for Na-Ion Batteries via Charged-Only Structures and Committee Divide-and-Conquer Learning
AU - Suchona Akter
AU - Md Omar Faruque
AU - Mohammad R. Momeni
PY - 2026
DA - 2026/10/03
PB - arXiv
DO - 10.48550/arxiv.2605.27229
UR - https://arxiv.org/abs/2605.27229
AB - Current cathode-active materials (CAMs) designed for sodium-ion batteries (SIBs) suffer from low operating voltages, limited energy density, irreversible phase transitions, and low intrinsic electrical conductivity. In this work, we have developed generalizable and transferable descriptor-based machine learning (ML) models, along with composition- and graph-based neural networks (NNs) and transfer learning (TL) models, all trained on a dataset of charged-only structures rather than charged/discharged pairs. Using a committee of the top four trained ML models, the mean voltage is predicted as the target property. More importantly, we show that uncertainty quantification can improve the reliability of candidate selection for training subsequent higher-generation models by including structures for which the current committee models are least confident. In this way, the new expensive ab initio data provide information in regions where the original models most need improvement. We have showcased the transferability of the final generation models by applying them to a large database of Na-free inorganic host structures curated from Materials Project, AFLOW, OQMD, and GNoME for at-scale SIB CAM discovery. A comprehensive interactive online database, NaCAM-Atlas, was also developed, containing 14,843 CAMs with voltages greater than 3 V for SIBs. Finally, a subset of the top-ranked candidate CAMs was validated through explicit, high-throughput first-principles calculations of voltage profiles, structural robustness upon sodiation/desodiation, and electronic properties. Together, this integrated data curation, ML training and improvement, and first-principles validation strategy establishes a scalable and transferable framework for accelerating the discovery of stable, high-voltage CAMs for SIBs and beyond.
ER -