Generation of magnetic metal-organic frameworks

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.

• The paper presents a novel generative approach using a site substitution protocol integrated with fine-tuned ML potentials to optimize magnetic properties in MOFs.
• The methodology leverages an extensive organic dataset to significantly reduce energy, force, and magnetic moment prediction errors compared to base models.
• The practical workflow screened over 149,000 candidates, ultimately validating 16 dynamically stable MOFs through DFT and phonon stability analyses.

Generation of Magnetic Metal-Organic Frameworks: A Technical Assessment

Problem Context

The design and discovery of magnetic metal-organic frameworks (MOFs) pose significant computational and methodological challenges, primarily due to their large unit cells and compositional complexity. Traditional first-principles methods like DFT scale cubically with atom count, rendering high-throughput screening of MOFs for advanced applications impractical. Moreover, the lack of organic material representation in foundational machine learning potential (MLP) training sets further impedes the use of efficient ML-driven approaches in this domain. This work systematically addresses these obstacles by curating a large organic-focused dataset and deploying fine-tuned MLPs for generative material design targeting highly magnetic MOFs.

Dataset Construction and Model Fine-Tuning

The authors utilize the Organic Materials Database (OMDB) to extract 15,000 single-point DFT calculations, explicitly including compounds with annotated magnetic moments and near-equilibrium structures. Fine-tuning is performed on CHGNet v0.3.0, a universal MLP distinguished by its capacity to model magnetic moments. The optimization protocol employs a 50-epoch cosine annealing learning rate schedule and assigns a force-driven weighted MAE loss to account for structural flexibility inherent in MOFs.

Numerical performance metrics demonstrate substantial improvements post fine-tuning:

• Energy prediction MAE: Reduced from 0.06 eV (base) to 0.02 eV (fine-tuned)
• Force prediction MAE: Reduced from 0.097 eV/Å (base) to 0.081 eV/Å (fine-tuned)
• Magnetic moment prediction MAE: Reduced from 0.010 μB (base) to 0.001 μB (fine-tuned)

These results signify a robust enhancement in CHGNet's performance on organic, magnetic materials, validating the efficacy of the OMDB-Traj dataset for ML-driven atomistic modeling.

Generative Workflow and Material Discovery

Rather than relying on deep generative neural architectures (diffusion, adversarial, or LLM-based models) -- which struggle with symmetry and chemical stability in large MOFs -- the authors introduce a systematic "site substitution" protocol. Magnetic MOF prototypes from the QMOF database serve as structural templates. All inequivalent atomic sites hosting magnetic elements (V, Cr, Mn, Fe, Co, Ni) are considered for exhaustive permutation and replacement with other magnetic atoms. Structural relaxation and magnetic property prediction are executed via the fine-tuned CHGNet.

The workflow achieves remarkable throughput:

• Generated candidate structures: >149,000
• Selection: Top 100 MOFs with highest predicted magnetic moments
• Validation: DFT calculations and phonon stability analysis reduce candidates to 16 dynamically stable compounds

Results and Technical Insights

The validated MOFs display a diverse spread in total magnetic moment amplitudes. Notably, site substitution is shown to be effective not only for enhancing magnetic properties but also for improving stability and, potentially, synthesizability. It is explicitly demonstrated that increases in magnetic moment are not guaranteed by all substitutions; this is attributed to the complex interplay of symmetry, site equivalence, and ligand-field effects.

For select compounds, DFT calculations confirm substantial magnetic moments, with one example (low-symmetry structure) exhibiting an increased total magnetic moment relative to its parent structure. Phonon and MD simulations further corroborate the dynamic stability of the final materials, aligning computational predictions with experimental feasibility.

Practical and Theoretical Implications

The paper provides a scalable, efficient methodology for generative design of magnetic MOFs, circumventing the limitations of both traditional DFT and generative deep learning models in the organic materials space. The approach enables:

• Rapid screening and optimization of large numbers of magnetic MOFs, fostering exploration of magnetic alternatives to rare-earth compounds.
• Integration of magnetic property prediction directly into generative frameworks, relevant for spintronic and magnetocaloric application domains.
• Potential extension of the site substitution workflow to other property targets (e.g., conductivity, catalytic activity, porosity), given the generalizability of fine-tuned MLPs in OMDB-contexts.

Theoretically, this research advances the ML-driven inverse design paradigm in materials science by providing a data-rich, organic-focused foundation for electronic and magnetic property optimization.

Speculation on Future Developments

As organic material databases grow and foundational MLPs mature, further fine-tuning and model extension will likely enable generative workflows for multi-property optimization, including stability, magnetism, and electronic transport. Feedback loops integrating experimental synthesis and ML-driven discovery are expected to accelerate. Additionally, hybrid strategies combining site substitution with latent-space generative models may address synthesis constraints while maintaining novelty and high-throughput capacity.

Conclusion

This study advances both methodological and computational frontiers in magnetic MOF discovery by leveraging curated organic datasets and fine-tuned machine-learned potentials. The presented workflow achieves high-throughput generation and accurate prediction of magnetic moments and stability for thousands of candidate MOFs, ultimately validating a select set of materials via first-principles calculations. The results facilitate practical exploration of MOFs as rare-earth alternatives, laying the groundwork for future ML-assisted material design innovations targeting a multitude of electronic and magnetic functionalities.

(2604.27879)