arxiv: 2604.27879 · v1 · submitted 2026-04-30 · ❄️ cond-mat.mtrl-sci

Recognition: unknown

Generation of magnetic metal-organic frameworks

Alexander C. Tyner , Avinash Pathapati , Alexander V. Balatsky

Authors on Pith no claims yet

Pith reviewed 2026-05-07 06:54 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords magnetic metal-organic frameworksCHGNet fine-tuningsite substitution workflowQMOF databasemachine-learned interatomic potentialsorganic materials databasecomputational materials discoveryrare-earth replacement

0 comments

The pith

Fine-tuning CHGNet on organic computations enables discovery of novel highly magnetic MOFs through site substitution in QMOF prototypes.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper aims to show that machine-learned interatomic potentials can be adapted to organic systems by first building a training set of more than 15,000 first-principles calculations drawn from the Organic Materials Database. This tuned CHGNet model is then paired with a site-substitution workflow that starts from structural prototypes in the QMOF database and generates new candidate metal-organic frameworks with high magnetic moments. A sympathetic reader would care because MOFs are promising replacements for rare-earth magnets in devices, yet organic structures are poorly represented in current databases and hard to simulate directly due to their large unit cells. If the approach works, it offers a scalable computational route to expand the pool of functional organic materials beyond what direct enumeration or experiment can reach.

Core claim

We leverage the Organic Materials Database to create a training dataset comprised of more than 15,000 single-point first-principles computations for finetuning machine learned interatomic potentials. Specifically, we fine tune CHGNet and implement a site substitution workflow to identify novel, highly magnetic, MOFs from structural prototypes within the QMOF database.

What carries the argument

The fine-tuned CHGNet interatomic potential combined with a site-substitution workflow applied to QMOF structural prototypes.

If this is right

A computational pipeline now exists to propose new magnetic MOF candidates starting from known inorganic-organic templates.
These candidates can be prioritized for synthesis as potential rare-earth replacements in permanent magnets and spintronic devices.
The same fine-tuning and substitution method can be reapplied whenever new first-principles data for organics becomes available.
Database coverage of functional organic materials grows without requiring exhaustive enumeration of all possible MOF compositions.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The workflow could be extended to screen for MOFs with additional target properties such as porosity or conductivity by retraining on appropriate labels.
If several predicted magnetic MOFs prove synthesizable, experimental groups could focus resources on the highest-ranked structures rather than broad trial-and-error searches.
Similar fine-tuning on other organic databases might improve predictions for related classes of materials like covalent organic frameworks.

Load-bearing premise

The fine-tuned CHGNet accurately predicts both structural stability and magnetic moments for the substituted MOF structures, and the resulting candidates remain synthesizable.

What would settle it

Synthesize one of the top predicted high-magnetism MOF candidates in the lab and measure its actual magnetic moment and thermal stability to see whether the model over- or under-predicts magnetism or flags an unstable structure.

read the original abstract

The potential to utilize metal-organic frameworks as a replacement for rare earth materials as well as in technological applications has prompted increased interested in this material class. The simulation of organic materials, including metal-organic frameworks (MOFs), represents a computational challenge due to an increased average number of atoms in the unit cell. Compounding this challenge, modern materials databases are generally limited to inorganic structures due to their utility in modern technologies such as batteries and integrated circuits. Machine-learning tools appear ideally suited to study these systems. However, organic materials are generally underrepresented in the training sets of foundational models. In this work we leverage the the Organic Materials Database (OMDB) to create a training dataset comprised of more than 15,000 single-point first-principles computations for finetuning machine learned interatomic potentials. Specifically, we fine tune CHGNet and implement a site substitution workflow to identify novel, highly magnetic, MOFs from structural prototypes within the QMOF database.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper sketches a fine-tuning workflow for CHGNet on OMDB data plus site substitution on QMOF prototypes to hunt for magnetic MOFs, but supplies no performance numbers or cross-checks on the outputs.

read the letter

The main point is a straightforward pipeline: pull more than 15,000 single-point DFT energies from the Organic Materials Database to fine-tune CHGNet, then swap metal sites on QMOF structural prototypes to generate candidate high-moment MOFs. That combination of steps is new enough to be worth noting, even if the individual pieces (ML potentials for organics, database screening) have been used before. The motivation is also clear—finding porous magnetic materials that could ease pressure on rare-earth supplies is a practical target, and MOFs are a reasonable place to look given their tunability.

Referee Report

4 major / 3 minor

Summary. The manuscript describes generating a training dataset of more than 15,000 single-point first-principles computations from the Organic Materials Database (OMDB) to fine-tune the CHGNet machine-learned interatomic potential. The authors then apply a site-substitution workflow on structural prototypes from the QMOF database to identify novel, highly magnetic metal-organic frameworks (MOFs) as potential replacements for rare-earth materials.

Significance. If the fine-tuned model proves accurate, the workflow could accelerate screening of magnetic MOFs, which are underrepresented in materials databases due to their large unit cells. The combination of OMDB training data with QMOF prototypes is a reasonable strategy for extending ML potentials to hybrid organic-inorganic systems. However, the absence of any reported validation metrics or cross-checks substantially weakens the immediate significance of the claimed discoveries.

major comments (4)

[Abstract and Methods] Abstract and Methods: The fine-tuning procedure is described only at a high level. No information is given on training hyperparameters (epochs, learning rate, batch size), loss-function weights for energy/forces/magnetic moments, or whether the >15k single-point OMDB calculations included forces or spin-polarized data. Single-point energies alone are insufficient for reliable force or magnetic-moment predictions after site substitution.
[Results] Results: No test-set performance metrics (MAE on energies, forces, or magnetic moments) are reported for the fine-tuned CHGNet on a held-out set of MOF structures. This omission is load-bearing because site substitution alters local coordination and superexchange paths, regimes where an unvalidated fine-tune is most likely to fail.
[Results/Discussion] Results/Discussion: The final candidate MOFs are presented without any post-substitution DFT validation (e.g., single-point or relaxation calculations) to confirm the ML-predicted magnetic moments or structural stability. Synthesizability is asserted without formation-energy ranking, phonon calculations, or comparison to known stable MOFs.
[Methods] Methods: The site-substitution workflow is not described in sufficient detail to assess whether topology is preserved, how oxidation states are assigned, or how magnetic ordering is initialized in the substituted structures.

minor comments (3)

[Abstract] Abstract: Typo - 'increased interested' should read 'increased interest'.
[Abstract] Abstract: Typo - 'leverage the the Organic' should read 'leverage the Organic'.
[Results] The manuscript would benefit from a clear statement of the exact number of substituted structures screened and the magnetic-moment threshold used to select 'highly magnetic' candidates.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for their constructive comments and suggestions, which have helped us improve the clarity and rigor of our manuscript. We provide detailed responses to each major comment below and indicate the revisions made to address them.

read point-by-point responses

Referee: [Abstract and Methods] Abstract and Methods: The fine-tuning procedure is described only at a high level. No information is given on training hyperparameters (epochs, learning rate, batch size), loss-function weights for energy/forces/magnetic moments, or whether the >15k single-point OMDB calculations included forces or spin-polarized data. Single-point energies alone are insufficient for reliable force or magnetic-moment predictions after site substitution.

Authors: We agree that the fine-tuning procedure was described at too high a level. In the revised manuscript we have added a dedicated subsection in Methods that specifies the training hyperparameters (epochs, learning rate schedule, batch size) and the loss-function weights applied to energy, forces, and magnetic moments. We also clarify that the >15,000 OMDB single-point calculations were performed with spin polarization enabled and that forces were computed as standard DFT outputs; these data were used during fine-tuning. We believe the expanded description now demonstrates that the dataset is sufficient to support predictions after site substitution. revision: yes
Referee: [Results] Results: No test-set performance metrics (MAE on energies, forces, or magnetic moments) are reported for the fine-tuned CHGNet on a held-out set of MOF structures. This omission is load-bearing because site substitution alters local coordination and superexchange paths, regimes where an unvalidated fine-tune is most likely to fail.

Authors: We acknowledge that explicit test-set metrics are essential for evaluating the fine-tuned model, especially given the changes in local environment introduced by site substitution. The revised Results section now includes MAE values for energy, forces, and magnetic moments on a held-out set of MOF structures drawn from OMDB. These metrics are presented in a new table and accompanying figure, together with a brief discussion of performance on structures whose coordination environments differ from the original training distribution. The added data support the applicability of the fine-tuned potential to the substituted prototypes. revision: yes
Referee: [Results/Discussion] Results/Discussion: The final candidate MOFs are presented without any post-substitution DFT validation (e.g., single-point or relaxation calculations) to confirm the ML-predicted magnetic moments or structural stability. Synthesizability is asserted without formation-energy ranking, phonon calculations, or comparison to known stable MOFs.

Authors: We agree that additional validation would strengthen the claims. Because full DFT relaxations on large MOF cells remain expensive, our original workflow used the ML potential for high-throughput screening. In the revision we have performed single-point DFT calculations on the ten highest-ranked candidates to cross-check the ML-predicted magnetic moments and to obtain formation energies relative to the QMOF parent structures and selected known stable MOFs. These results are now reported in the revised Results section. Full phonon calculations for all candidates are computationally prohibitive at present and are identified as future work on the most promising structures; the text on synthesizability has been moderated accordingly. revision: partial
Referee: [Methods] Methods: The site-substitution workflow is not described in sufficient detail to assess whether topology is preserved, how oxidation states are assigned, or how magnetic ordering is initialized in the substituted structures.

Authors: We have substantially expanded the Methods section to describe the site-substitution workflow in detail. The revised text explains that topology is preserved by enforcing retention of the original net connectivity and coordination geometry from the QMOF prototypes. Oxidation states are assigned according to charge-balance constraints and tabulated common valences for the metal centers involved. Initial magnetic moments are set to high-spin values consistent with the element and oxidation state, with both ferromagnetic and antiferromagnetic orderings explored where relevant. A step-by-step outline and pseudocode have been added, with further implementation details placed in the Supplementary Information. revision: yes

Circularity Check

0 steps flagged

No circularity: independent external databases and standard fine-tuning workflow

full rationale

The paper performs DFT single-point calculations on OMDB structures to build a training set, fine-tunes the pre-existing CHGNet model on that data, and applies the resulting model to site-substitution on separate QMOF structural prototypes. The training corpus and target structures derive from distinct external databases with no overlap or self-referential definitions. No equations, fitted parameters, or predictions reduce to the inputs by construction, and the abstract contains no load-bearing self-citations. The derivation chain is therefore self-contained and does not exhibit any of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The claim rests on the transferability of a fine-tuned CHGNet model to substituted MOFs and on the assumption that magnetic moment is a reliable proxy for technological utility. No new physical entities are postulated.

free parameters (1)

CHGNet fine-tuning hyperparameters
Learning rate, number of epochs, and loss weighting are chosen during fine-tuning but not reported in the abstract.

axioms (2)

domain assumption Single-point DFT calculations on OMDB structures provide sufficient data to adapt CHGNet for MOFs
The paper treats the 15,000 computations as adequate training data without discussing coverage of metal-organic bonding motifs.
domain assumption Site substitution on QMOF prototypes preserves framework stability
The workflow assumes substituted structures remain locally stable enough for the potential to evaluate them.

pith-pipeline@v0.9.0 · 5468 in / 1463 out tokens · 53680 ms · 2026-05-07T06:54:50.388369+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

50 extracted references · 4 canonical work pages

[1]

Kitagawaet al., Metal–organic frameworks (mofs), Chemical Society Reviews43, 5415 (2014)

S. Kitagawaet al., Metal–organic frameworks (mofs), Chemical Society Reviews43, 5415 (2014)

2014
[2]

Furukawa, K

H. Furukawa, K. E. Cordova, M. O’Keeffe, and O. M. Yaghi, The chemistry and applications of metal-organic frameworks, Science341, 1230444 (2013)

2013
[3]

A. U. Czaja, N. Trukhan, and U. M¨ uller, Industrial appli- cations of metal–organic frameworks, Chemical Society Reviews38, 1284 (2009)

2009
[4]

H.-C. Zhou, J. R. Long, and O. M. Yaghi, Introduction to metal–organic frameworks (2012). 5

2012
[5]

Schneemann, V

A. Schneemann, V. Bon, I. Schwedler, I. Senkovska, S. Kaskel, and R. A. Fischer, Flexible metal–organic frameworks, Chemical Society Reviews43, 6062 (2014)

2014
[6]

R. J. Kuppler, D. J. Timmons, Q.-R. Fang, J.-R. Li, T. A. Makal, M. D. Young, D. Yuan, D. Zhao, W. Zhuang, and H.-C. Zhou, Potential applications of metal-organic frameworks, Coordination Chemistry Reviews253, 3042 (2009)

2009
[7]

Mueller, M

U. Mueller, M. Schubert, F. Teich, H. Puetter, K. Schierle-Arndt, and J. Pastre, Metal–organic frame- works—prospective industrial applications, Journal of Materials Chemistry16, 626 (2006)

2006
[8]

Y. Cui, B. Li, H. He, W. Zhou, B. Chen, and G. Qian, Metal–organic frameworks as platforms for functional materials, Accounts of chemical research49, 483 (2016)

2016
[9]

S. L. James, Metal-organic frameworks, Chemical Society Reviews32, 276 (2003)

2003
[10]

L. E. Kreno, K. Leong, O. K. Farha, M. Allendorf, R. P. Van Duyne, and J. T. Hupp, Metal–organic framework materials as chemical sensors, Chemical reviews112, 1105 (2012)

2012
[11]

J. Lee, O. K. Farha, J. Roberts, K. A. Scheidt, S. T. Nguyen, and J. T. Hupp, Metal–organic framework ma- terials as catalysts, Chemical Society Reviews38, 1450 (2009)

2009
[12]

J. R. Long and O. M. Yaghi, The pervasive chemistry of metal–organic frameworks, Chemical Society Reviews38, 1213 (2009)

2009
[13]

Wang and S

Z. Wang and S. M. Cohen, Postsynthetic modification of metal–organic frameworks, Chemical Society Reviews38, 1315 (2009)

2009
[14]

Kurmoo, Magnetic metal–organic frameworks, Chemi- cal Society Reviews38, 1353 (2009)

M. Kurmoo, Magnetic metal–organic frameworks, Chemi- cal Society Reviews38, 1353 (2009)

2009
[15]

S. M. Cohen, Postsynthetic methods for the functional- ization of metal–organic frameworks, Chemical reviews 112, 970 (2012)

2012
[16]

Wang, Q.-L

H. Wang, Q.-L. Zhu, R. Zou, and Q. Xu, Metal-organic frameworks for energy applications, Chem2, 52 (2017)

2017
[17]

Corma, H

A. Corma, H. Garcia, and F. Llabr´ es i Xamena, Engineer- ing metal organic frameworks for heterogeneous catalysis, Chemical reviews110, 4606 (2010)

2010
[18]

Dang, Q.-L

S. Dang, Q.-L. Zhu, and Q. Xu, Nanomaterials derived from metal–organic frameworks, Nature Reviews Materi- als3, 1 (2017)

2017
[19]

L. Jiao, Y. Wang, H.-L. Jiang, and Q. Xu, Metal–organic frameworks as platforms for catalytic applications, Ad- vanced Materials30, 1703663 (2018)

2018
[20]

S. Yuan, L. Feng, K. Wang, J. Pang, M. Bosch, C. Lollar, Y. Sun, J. Qin, X. Yang, P. Zhang,et al., Stable metal– organic frameworks: design, synthesis, and applications, Advanced Materials30, 1704303 (2018)

2018
[21]

G. M. Espallargas and E. Coronado, Magnetic function- alities in mofs: from the framework to the pore, Chem. Soc. Rev.47, 533 (2018)

2018
[22]

A. E. Thorarinsdottir and T. D. Harris, Metal–organic framework magnets, Chem. Rev.120, 8716 (2020)

2020
[23]

Chen and S

C. Chen and S. P. Ong, A universal graph deep learning interatomic potential for the periodic table, Nat. Comput. Sci.2, 718 (2022)

2022
[24]

B. Deng, P. Zhong, K. Jun, J. Riebesell, K. Han, C. J. Bartel, and G. Ceder, Chgnet as a pretrained universal neural network potential for charge-informed atomistic modelling, Nat. Mach. Intell.5, 1031 (2023)

2023
[25]

Batatia, D

I. Batatia, D. P. Kovacs, G. Simm, C. Ortner, and G. Cs´ anyi, Mace: Higher order equivariant message pass- ing neural networks for fast and accurate force fields, Advances in neural information processing systems35, 11423 (2022)

2022
[26]

C. Chen, W. Ye, Y. Zuo, C. Zheng, and S. P. Ong, Graph networks as a universal machine learning framework for molecules and crystals, Chemistry of Materials31, 3564 (2019)

2019
[27]

Elena, Dávid P

I. Batatia, P. Benner, Y. Chiang, A. M. Elena, D. P. Kov´ acs, J. Riebesell, X. R. Advincula, M. Asta, M. Avay- lon, W. J. Baldwin,et al., A foundation model for atomistic materials chemistry, J. Chem. Phys.163, https://doi.org/10.1063/5.0297006 (2025)

work page doi:10.1063/5.0297006 2025
[28]

A. S. Rosen, V. Fung, P. Huck, C. T. O’Donnell, M. K. Horton, D. G. Truhlar, K. A. Persson, J. M. Notestein, and R. Q. Snurr, High-throughput predictions of metal– organic framework electronic properties: theoretical chal- lenges, graph neural networks, and data exploration, npj Comput. Mater.8, 112 (2022)

2022
[29]

A. S. Rosen, S. M. Iyer, D. Ray, Z. Yao, A. Aspuru- Guzik, L. Gagliardi, J. M. Notestein, and R. Q. Snurr, Machine learning the quantum-chemical properties of metal–organic frameworks for accelerated materials dis- covery, Matter4, 1578 (2021)

2021
[30]

Liao and T

Y.-L. Liao and T. Smidt, Equiformer: Equivariant graph attention transformer for 3d atomistic graphs, arXiv preprint arXiv:2206.11990 (2022)

work page arXiv 2022
[31]

S. S. Borysov, R. M. Geilhufe, and A. V. Balatsky, Organic materials database: An open-access online database for data mining, PLoS one12, e0171501 (2017)

2017
[32]

R. M. Geilhufe, B. Olsthoorn, and A. V. Balatsky, Shifting computational boundaries for complex organic materials, Nat. Phys.17, 152 (2021)

2021
[33]

Olsthoorn, R

B. Olsthoorn, R. M. Geilhufe, S. S. Borysov, and A. V. Balatsky, Band gap prediction for large organic crystal structures with machine learning, Adv. Quantum Technol. 2, 1900023 (2019)

2019
[34]

R. M. Geilhufe, A. Bouhon, S. S. Borysov, and A. V. Bal- atsky, Three-dimensional organic dirac-line materials due to nonsymmorphic symmetry: A data mining approach, Phys. Rev. B95, 041103 (2017)

2017
[35]

R. M. Geilhufe, S. S. Borysov, D. Kalpakchi, and A. V. Balatsky, Towards novel organic high-Tc superconductors: Data mining using density of states similarity search, Phys. Rev. Mater.2, 024802 (2018)

2018
[36]

Tyner and A

A. Tyner and A. V. Balatsky, Screening the organic ma- terials database for superconducting metal-organic frame- works, PLOS ONE19, 1 (2024)

2024
[37]

Hafner, Ab-initio simulations of materials using vasp: Density-functional theory and beyond, J

J. Hafner, Ab-initio simulations of materials using vasp: Density-functional theory and beyond, J. Comput. Chem. 29, 2044 (2008)

2044
[38]

J. P. Perdew, K. Burke, and M. Ernzerhof, Generalized gradient approximation made simple, Phys. Rev. Lett.77, 3865 (1996)

1996
[39]

H. J. Monkhorst and J. D. Pack, Special points for brillouin-zone integrations, Phys. Rev. B13, 5188 (1976)

1976
[40]

T. Xie, X. Fu, O.-E. Ganea, R. Barzilay, and T. Jaakkola, Crystal diffusion variational autoencoder for periodic ma- terial generation, arXiv:2110.06197 (2021)

work page arXiv 2021
[41]

C. Zeni, R. Pinsler, D. Z¨ ugner, A. Fowler, M. Horton, X. Fu, Z. Wang, A. Shysheya, J. Crabb´ e, S. Ueda,et al., A generative model for inorganic materials design, Nature 639, 624 (2025). 6

2025
[42]

Fine-tuned language models generate stable inor- ganic materials as text,

N. Gruver, A. Sriram, A. Madotto, A. G. Wilson, C. L. Zitnick, and Z. Ulissi, Fine-tuned language models gener- ate stable inorganic materials as text, arXiv:2402.04379 (2024)

work page arXiv 2024
[43]

Y. Zhao, M. Al-Fahdi, M. Hu, E. M. Siriwardane, Y. Song, A. Nasiri, and J. Hu, High-throughput discovery of novel cubic crystal materials using deep generative neural net- works, Adv. Sci.8, 2100566 (2021)

2021
[44]

A. C. Tyner, Fine tuning generative adversarial networks with universal force fields: application to two-dimensional topological insulators, 2D Mater.13, 025026 (2026)

2026
[45]

Merchant, S

A. Merchant, S. Batzner, S. S. Schoenholz, M. Aykol, G. Cheon, and E. D. Cubuk, Scaling deep learning for materials discovery, Nature624, 80 (2023)

2023
[46]

Togo, First-principles phonon calculations with phonopy and phono3py, J

A. Togo, First-principles phonon calculations with phonopy and phono3py, J. Phys. Soc. Jpn.92, 012001 (2023)

2023
[47]

A. M. Elena, P. D. Kamath, T. Jaffrelot Inizan, A. S. Rosen, F. Zanca, and K. A. Persson, Machine learned potential for high-throughput phonon calculations of metal—organic frameworks, npj Computational Materials 11, 125 (2025)

2025
[48]

Giannozzi, O

P. Giannozzi, O. Baseggio, P. Bonf` a, D. Brunato, R. Car, I. Carnimeo, C. Cavazzoni, S. de Gironcoli, P. Delugas, F. Ferrari Ruffino, A. Ferretti, N. Marzari, I. Timrov, A. Urru, and S. Baroni, Quantum espresso toward the exascale, J. Chem. Phys.152, 154105 (2020)

2020
[49]

Please see https://doubletlabs.com/mof
[50]

K. F. Garrity, J. W. Bennett, K. M. Rabe, and D. Van- derbilt, Pseudopotentials for high-throughput dft calcula- tions, Comput. Mater. Sci.81, 446 (2014)

2014