Machine Learning Identifies Operating Boundaries for Organic Framework Battery Electrodes

Sustainable organic electrode materials have emerged as compelling alternatives to traditional inorganic battery components that often suffer from resource scarcity, high economic costs, energy-intensive processing, and severe environmental concerns. Among the diverse architectures explored for rechargeable energy storage, two-dimensional covalent organic frameworks, commonly termed 2D COFs, have attracted substantial investigation due to their ordered porous polymeric networks constructed from lightweight elements joined by robust covalent bonds. These crystalline structures feature reticular design principles, high surface areas, and abundantly accessible pore channels that allow researchers to deliberately incorporate redox-active functionalities while tailoring local pore environments for efficient ion transport and adsorption. Favorable molecular stacking, extended conjugation pathways, and the introduction of conductive additives have further improved the electronic conductivity and cycling stability of these porous networks in preliminary laboratory trials. Despite extensive experimental investigations across the past decade, the design of electroactive framework electrodes has remained overwhelmingly empirical because the fundamental structure-performance relationships governing their operational behavior are poorly understood, leaving their true viability as functional battery electrodes unresolved.

To establish a systematic understanding of these materials, researchers developed a data-driven computational framework combining density functional theory computations and machine learning models to map structure-electrode relationships across thousands of synthesized two-dimensional frameworks. A viable metal-ion battery electrode material must deliver an appropriate operating voltage and reversible capacity, enable sufficiently rapid ion and electron transport, and maintain electrochemical, thermal, and mechanical stability across repeated charge and discharge cycling. Among these primary thermodynamic metrics, the open-circuit voltage referenced against lithium metal serves as the decisive descriptor that differentiates potential anodic and cathodic behavior. Because covalent organic frameworks naturally lack rigidly defined operational voltage windows, an analysis of experimentally reported electrodes identified approximately 1.6 volts versus lithium as an appropriate classification threshold. Materials operating below 1.6 volts function as anodes by undergoing reduction at lower potentials, whereas frameworks exhibiting potentials at or above 1.6 volts are positioned as cathodes, matching the reduction potential windows previously established for organic redox functionalities.

Assessing candidate structures through purely experimental trial-and-error synthesis is impractical given the thousands of synthesized frameworks reported to date, while direct brute-force quantum mechanical calculations across massive chemical libraries remain computationally prohibitive. To overcome these screening limitations, the investigators curated a chemically diverse training library from the Computation-Ready Experimental Covalent Organic Framework database, which contains 1099 crystal structures derived from synthesized materials. After removing duplicate entries, disordered structures, and metal-containing networks, the team assembled an updated list of 996 candidate structures and selected a representative subset of 140 diverse frameworks designated as the Diverse Inventory of CoRE COFs for Electrodes. This curated collection spans the primary redox-active functionalities explored in organic battery research, incorporating carbonyl systems such as quinones, ketones, and imides, imine architectures including azines, hydrazones, and phenazines, nitro groups, macrocyclic porphyrins and phthalocyanines, and electron-deficient triazine and heptazine ring systems to capture a comprehensive spectrum of metal-ion binding thermodynamics.

For each selected framework within this curated inventory, the researchers calculated theoretical open-circuit voltages and gravimetric storage capacities using first-principles density functional theory based on the generalized gradient approximation with dispersion corrections. Taking representative systems such as Tb-DANT-COF, the automated workflow sequentially determined lithium-ion binding energies across all crystallographically unique adsorption sites, initiating lithiation at the most energetically favorable position before systematically occupying remaining accessible locations until further insertion became thermodynamically unfavorable. This computational evaluation produced two benchmarked training sets, designated DICE-V for voltage classification and DICE-C for capacity prediction, each comprising 137 distinct framework structures with 134 shared entries. Dimensionality reduction via t-distributed stochastic neighbor embedding verified that these curated subsets broadly represented the chemical and geometric diversity of the parent database without introducing domain bias, while applicability domain calculations confirmed that 94.7 percent of the remaining framework database fell within the descriptor space defined by the curated training data.

To convert structural geometries into machine learning inputs, the researchers extracted thirty-one composition-based and local geometric features from unrelaxed crystallographic information files for the voltage classification task. Six elemental composition ratios, the total degree of unsaturation, and total counts of electronegative atoms were combined with twenty-three localized descriptors characterizing coordination environments around sp2-hybridized oxygen and nitrogen atoms, proximal heteroatom distributions, and electronegativity-weighted radial distribution functions. Five additional unit cell parameters, including dimensions, molecular weight, and total active site counts, were incorporated for capacity modeling. Evaluating multiple predictive architectures revealed that an ensemble of five XGBoost classifiers utilizing recursive feature elimination achieved the highest classification accuracy. Across repeated five-by-five stratified cross-validation on the imbalanced dataset containing 88 percent anode-like and 12 percent cathode-like frameworks, the consensus voting voltage model demonstrated robust discriminative performance, yielding an area under the receiver operating characteristic curve of 0.88, a balanced accuracy of 0.77, and an out-of-fold precision-recall curve area of 0.51.

Feature importance analyses revealed that localized chemical environments around redox-active sites governed open-circuit voltage classification, contributing roughly 77 percent of total model importance, with proximal heteroatom parameters accounting for 65 percent alone. Frameworks exhibited cathode-like behavior when coordination pockets provided three or more coordination contacts to each bound lithium ion through proximate nitrogen and oxygen atoms. These dense multidentate environments typically arranged two donor sites in-plane at distances of 2.5 to 2.8 angstroms, while a third coordinating donor emerged from interlayer periodicity across 3.2 to 6.7 angstroms or suitably oriented side chains. Motifs promoting higher operating voltages included acylhydrazone linkages, functionalized keto-amine groups augmented with electron-withdrawing substituents, cyclic imide cores, nitrogen-rich hexaazatriphenylene rings offering multi-lithium coordination, and conjugated porphyrin rings providing preorganized four-nitrogen chelation pockets. Conversely, compositional features such as overall nitrogen and oxygen ratios contributed only 23 percent of importance, confirming that localized spatial coordination outweighs broad bulk composition when establishing redox potentials.

Applying the trained ensemble models across the complete database demonstrated that 85 percent of synthesized two-dimensional frameworks operate as anode-like materials, while only 15 percent exhibit cathode-like characteristics. Anodic behavior correlated strongly with architectures featuring spatially separated redox motifs, such as isolated imine, triazine, pyridine, or five-membered heteroaromatic rings, which maintain shallow, moderate lithium binding that facilitates reversible deintercalation at low potentials without strong chemisorption. Additionally, lithium insertion within pi-rich interlayer galleries of extended pyrene, phenyl, and benzothiophene surfaces supported reversible uptake with characteristic interlayer spacings of approximately 3.7 angstroms. Meanwhile, capacity classification models, which achieved an area under the receiver operating characteristic curve of 0.75, predicted that 75 percent of candidates exceed 200 milliampere-hours per gram. This accessible capacity was governed predominantly by unit-cell structural dimensions, particularly the c-axis parameter that defines interlayer gallery volume, whereas boronate ester, boroxine, flexible aliphatic ether chains, and triarylamine units were identified as intrinsically poor storage motifs that provide minimal reversible lithium capacity.

To isolate high-performing candidates suitable for practical battery operations, the investigators implemented a multi-stage screening protocol combining rigorous thermodynamic thresholds with electronic structure evaluations. From an initial pool of 164 predicted candidates, stricter voltage criteria requiring open-circuit potentials below 0.62 volts for anodes and above 2.0 volts for cathodes narrowed the collection to 17 anode and 14 cathode frameworks. Because pristine organic frameworks are typically semiconducting, the researchers evaluated density of states at the HSE06 hybrid functional level to identify systems undergoing semiconductor-to-metal transitions upon initial lithiation, which ensures the high electronic conductivity vital for rapid charge transport and battery rate performance. This electronic screening confirmed that 28 lead framework candidates, comprising 14 anodes and 14 cathodes, intrinsically undergo metallization upon initial lithium insertion, validating theoretical predictions against established experimental cathode benchmarks including Tp-DANT-COF, Tb-DANT-COF, and DABQ-TFP-COF, which operate with experimental voltage plateaus between 2.5 and 2.8 volts.

Direct performance comparisons between these 28 optimized organic candidates and conventional commercial electrodes highlighted critical technological trade-offs for modern energy storage devices. The screened framework anodes delivered theoretical operating voltages between 0.26 and 0.62 volts versus lithium and reversible capacities spanning 200 to 577 milliampere-hours per gram. Operating above the 0.1 to 0.2 volt plateau of standard graphite anodes enables these framework materials to significantly mitigate dangerous lithium plating risks during fast-charging or low-temperature regimes, with several candidates exceeding graphite’s theoretical capacity of 372 milliampere-hours per gram. Although framework anodes underperform compared to silicon anodes that reach theoretical capacities of roughly 3579 milliampere-hours per gram, their rigid porous channels may accommodate ion insertion without the catastrophic 300 percent volumetric expansion characteristic of silicon lithiation. Conversely, the screened organic cathodes exhibited operating voltages falling one to two volts below prevalent inorganic commercial cathodes like lithium cobalt oxide, lithium iron phosphate, and nickel manganese cobalt oxides, because organic nitrogen and oxygen functionalities possess weaker inherent redox potentials than electropositive transition-metal redox couples.

These operational voltage limitations indicate that two-dimensional covalent organic frameworks cannot serve as straightforward drop-in replacements for standard inorganic cathodes in conventional lithium-ion batteries. However, this thermodynamic reality shifts dramatically when pairing optimized organic frameworks together within rocking-chair all-organic lithium batteries. Systematic evaluation of 196 possible anode-cathode pairings identified 61 distinct combinations capable of delivering full-cell operational voltages at or above 2.0 volts, alongside full-cell capacities ranging from 101 to 290 milliampere-hours per gram and calculated energy densities reaching up to 575 watt-hours per kilogram. Beyond all-organic cells, the identified low-potential framework anodes provide chemically stable, dendrite-resistant alternatives for lithium-sulfur, lithium-selenium, and lithium-oxygen batteries, while framework cathodes require tailored chemical engineering to address complex dissolution and passivation reactions. Ultimately, this integrated machine learning and density functional theory approach establishes rational atomic-level guidelines that replace trial-and-error discovery, providing a clear blueprint for engineering next-generation organic frameworks across diverse metal-ion battery chemistries.

Subject of Research: Data-driven computational screening of 2D covalent organic framework electrodes for lithium-ion batteries

Article Title: When are Two‐Dimensional Covalent Organic Frameworks Good Electrodes for Lithium Ion Batteries? A Data‐Driven Analysis

Article References: Bhauriyal, P., Das, S., & Heine, T. (2026). When are Two‐Dimensional Covalent Organic Frameworks Good Electrodes for Lithium Ion Batteries? A Data‐Driven Analysis. Advanced Science, Article e78599. https://doi.org/10.1002/advs.78599

Image Credits: AI Generated

DOI: 10.1002/advs.78599

Keywords: covalent organic frameworks, lithium-ion batteries, density functional theory, machine learning, organic electrodes, energy storage, all-organic batteries, materials informatics, Dimensional, Covalent, Organic, Frameworks