TL;DR: University of Florida researchers published COFFlow on September 23, 2026, a flow-matching-based generative framework that assembles two-dimensional covalent organic frameworks from building-block representations and was used to identify previously unreported methane-storage candidates.

A new generative tool arrived on September 23

A team in the Department of Chemistry and the Quantum Theory Project at the University of Florida reported a new artificial-intelligence framework for porous-materials design, according to ACS Publications.

The paper, titled COFFlow: Building-Block-Based Generative Modeling for Two-Dimensional Covalent Organic Framework Discovery, was published online on September 23, 2026 in Chemistry of Materials.

The authors listed are Yuanhui Pan, Yunrui Yan, Cheng Zeng, Jirui Jin and Mingjie Liu, with Mingjie Liu identified as the corresponding author through the address [email protected], according to ACS Publications.

The article history records receipt on May 26, 2026, a revision received on September 11, 2026, and acceptance on September 14, 2026, before online publication on September 23, 2026.

The paper was assigned to volume 38, issue 19, pages 10036–10049, with publication in issue dated October 13, 2026, according to ACS Publications.

The publisher is the American Chemical Society, and the authors declared no competing financial interest.

Funding is listed as support from the University of Florida, with no separate award identifier shown.

The supporting record on ACS Figshare, posted under the same title and author list, carries the same abstract describing COFFlow and its methane-storage demonstration.

What two-dimensional COFs are

Two-dimensional covalent organic frameworks, abbreviated as 2D COFs, are described in the paper as promising porous materials, according to ACS Publications.

Their design space is described as vast yet largely unexplored, arising from diverse building blocks, abbreviated as BBs, and crystal topologies.

Covalent organic frameworks more broadly are a class of porous polymers that form two- or three-dimensional structures through reactions between organic precursors, resulting in strong covalent bonds that afford porous, stable and crystalline materials, according to the Covalent organic framework entry.

That entry describes COFs as reticular materials with tunable porosity and high surface area, attractive for applications including gas storage, catalysis, sensing and energy-related technologies.

Early 2D COFs identified as COF-1 and COF-5 were reported in 2005 from condensation of phenyl diboronic acid and hexahydroxytriphenylene, with powder X-ray diffraction showing expanded porous graphitic layers, according to the Covalent organic framework entry.

COF-1 was described with a staggered conformation and COF-5 with an eclipsed conformation, held by strong bonds between boron, carbon and oxygen atoms to form rigid porous architectures with pore sizes ranging from 7 to 27 angstroms, according to the same entry.

Reported thermal stability reached temperatures up to 500 to 600 degrees Celsius, with specific surface areas of 711 and 1590 square meters per gram for COF-1 and COF-5 respectively, according to the same entry.

Three-dimensional COFs designated COF-103 and COF-108 followed in work reported in 2007, extending the field beyond layered systems, according to the same entry.

Secondary building units, also called precursors, are conceptual fragments whose combinations can produce a very large number of frameworks with distinct properties for separations, storage and heterogeneous catalysis, according to the same entry.

Reticular synthesis refers to a bottom-up approach in which a material is built from atomic or molecular components synthetically, with control over composition enabling tuning of framework properties, according to the same entry.

Why prediction and inverse design remain difficult

Efficient crystal structure prediction, abbreviated as CSP, and inverse design of 2D COFs are described as highly challenging, according to ACS Publications.

The challenge is linked directly to the combination of diverse building blocks and diverse crystal topologies that defines the 2D COF space.

In that context, crystal structure prediction means determining a periodic crystal arrangement consistent with given chemical components, while inverse design means searching for structures intended to meet a target performance measure.

The paper positions both tasks as central barriers to exploring the largely unexplored portion of the 2D COF landscape.

The ACS Figshare record repeats the same characterization of CSP and inverse design as highly challenging for 2D COFs.

What COFFlow is

The team developed COFFlow, described as a flow-matching-based generative framework for CSP of 2D COFs, according to ACS Publications.

Flow matching in this paper denotes the generative modeling approach used to produce candidate crystal structures.

COFFlow is specifically directed to two-dimensional systems rather than to all porous polymers.

The framework is presented as enabling direct generative assembly of periodic frameworks from BB representations.

That assembly capability is tied to a COF-specific deconstruction algorithm introduced in the same work.

The paper states that systematic evaluations demonstrate that COFFlow can reliably and efficiently generate structurally consistent and chemically valid 2D COFs, according to ACS Publications.

Structural consistency here refers to agreement between generated periodic frameworks and expected crystal organization, while chemical validity refers to whether generated structures satisfy chemical rules applied by the authors.

The broader claim is that the work establishes a scalable framework for AI-driven discovery and design of reticular materials, according to ACS Publications.

Reticular materials in this usage include COFs alongside related framework families assembled from molecular units.

Deconstruction into experimentally relevant blocks

A central element is a COF-specific deconstruction algorithm, according to ACS Publications.

The algorithm is described as consistently decomposing periodic COFs into experimentally relevant BBs across diverse topologies.

Experimentally relevant in this description means building blocks connected to laboratory synthesis rather than arbitrary fragments.

Consistency across diverse topologies means the same decomposition logic applies when crystal connectivity changes.

The resulting BB representations then serve as the direct starting point for generative assembly of periodic frameworks.

The supporting information lists a dedicated section on the COF deconstruction algorithm, alongside data cleaning and preparation, according to ACS Publications.

That supporting document also lists model details, training strategies for the COFFlow models, and computational details for evaluation, indicating that deconstruction, data preparation and model training were documented together.

Examples and distributions of generated building blocks and classified topologies are listed among the additional supporting data, according to ACS Publications.

Samples of generated 2D COFs compared against ground-truth structures for each of the COFFlow models are also listed in the supporting information.

How results were measured: match rate and validity rate

Evaluation used the commonly used match rate, abbreviated as MR, according to ACS Publications.

Beyond MR, the authors introduced validity rate, abbreviated as VR, to assess the chemical validity of generated structures.

Match rate therefore represents an existing structural-consistency measure, while validity rate adds a separate chemical-validity measure.

The supporting information lists computational details for the evaluations of MR and VR, according to ACS Publications.

It also lists geometry property calculations and additional geometry property distributions, indicating that physical pore and lattice characteristics were examined alongside MR and VR.

Structural illustrations of representative generated structures and geometry properties of representative 2D COFs are listed as additional supporting data, according to ACS Publications.

The abstract’s evaluation summary states that COFFlow can reliably and efficiently generate structurally consistent and chemically valid 2D COFs under systematic evaluation, according to ACS Publications.

Tests beyond the training distribution

The models were tested for extrapolation toward unseen BBs and unseen crystal topologies, according to ACS Publications.

The paper describes that extrapolation capability as nontrivial, meaning performance extended beyond memorized training cases in a substantive way.

At the same time, the authors state that preserving chemical validity under strong distribution shifts remains challenging, according to ACS Publications.

A distribution shift in this context means evaluation on building blocks or topologies substantially different from those seen during training.

The finding therefore distinguishes structural extrapolation from maintenance of chemical validity when the shift is large.

The supporting information organizes supporting data into sections for a base model, a BB-focused model and a topology-focused model, according to ACS Publications.

Those section labels — COFFlow-base, COFFlow-BB and COFFlow-Topo — indicate separate evaluation of general performance, building-block shifts and topology shifts.

The supporting document also lists model-details sections covering flow matching on Riemannian manifolds, neural-network architecture associated with MOFFlow, and training strategies for the COFFlow models, according to ACS Publications.

Methane-storage demonstration

To demonstrate practical utility, the team integrated COFFlow with data-driven screening and inverse design of 2D COFs for methane storage, according to ACS Publications.

Deliverable capacity was used as the target metric for adsorbed natural gas application.

Adsorbed natural gas refers to storage of natural gas by adsorption in a porous material rather than by compression alone.

Using that metric, the authors reported identifying several previously unreported 2D COFs with performance among the highest reported to date, according to ACS Publications.

Previously unreported means the frameworks were not already present in the literature or databases considered by the authors.

Among the highest reported to date places those candidates near the top of performance values known to the authors at the time of writing.

Methane uptake results are listed among the supporting data, according to ACS Publications.

The supporting information also lists a section on screening for COFs with top deliverable capacity and a section on validation of the grand canonical Monte Carlo simulation protocol, abbreviated in the field as GCMC, according to ACS Publications.

That protocol validation section indicates that the adsorption-simulation method used for screening was checked separately.

Publication record and availability

The Chemistry of Materials record identifies the online ISSN as 1520-5002 and the print ISSN as 0897-4756, according to ACS Publications.

The digital object identifier is https://doi.org/10.1021/acs.chemmater.6c01524.

The supporting information is described as available free of charge at that DOI address, according to ACS Publications.

The supporting PDF is described as containing model details, data-preparation details, computational details for MR and VR, geometry-property calculations and additional supporting data, according to ACS Publications.

The listed additional data include geometry-property distributions, structural illustrations, geometry properties of representative 2D COFs, building-block and topology distributions, generated-versus-ground-truth samples and methane-uptake results.

The ACS Figshare item carrying the same title lists Yuanhui Pan, Yunrui Yan, Cheng Zeng, Jirui Jin and Mingjie Liu as authors and repeats the paper’s abstract.

The paper’s author affiliations place all five authors in both the Department of Chemistry and the Quantum Theory Project at the University of Florida in Gainesville, Florida 32611, United States, according to ACS Publications.

Why building-block generation matters for reticular chemistry

Design from building-block representations keeps generated frameworks closer to units that chemists recognize from synthesis, because the paper emphasizes experimentally relevant BBs, according to ACS Publications.

That emphasis matters for COFs because precursor selection and synthetic control were central to moving from amorphous organic polymers to porous crystalline materials, according to the Covalent organic framework entry.

Judicious selection of secondary building units allows the final structure to be predetermined and modified with control, enabling tuning of emergent properties, according to the same entry.

The COFFlow approach therefore connects generative modeling to reticular thinking, in which frameworks are understood as assemblies of defined units in defined topologies.

The claim of scalability for AI-driven discovery and design of reticular materials points beyond a single methane-storage test, according to ACS Publications.

At the same time, the reported difficulty of preserving chemical validity under strong distribution shifts bounds that claim, because unseen chemistry remains harder to generate validly.

The combination of MR and VR responds to that bound by separating structural agreement from chemical plausibility.

What remains tied to September 23

On the archive date of September 23, 2026, the established facts were publication, method description, evaluation framing and the methane-storage demonstration as stated in the abstract, according to ACS Publications and the ACS Figshare record.

Plans described as inverse design and data-driven screening were reported as integrated steps leading to candidate identification, not as later laboratory synthesis results, according to ACS Publications.

The candidates were described as previously unreported 2D COFs with deliverable-capacity performance among the highest reported to date, without additional performance figures in the abstract, according to ACS Publications.

Detailed distributions, structures, geometry properties and methane-uptake tables were assigned to supporting information rather than to the main abstract.

Future synthesis, experimental adsorption testing or deployment of those candidates were not stated in the abstract available by the cutoff.

The durable contribution documented by that date is therefore a building-block-based, flow-matching framework for 2D COF crystal structure prediction, a new validity-rate evaluation, evidence of extrapolation with stated limits, and a methane-storage design example.

Sources referenced