Phase 4 of our RE-free permanent magnet community building, with a strategic pivot.
Instead of broad cold outreach, this batch focuses on engaging computational materials scientists and ML researchers who have recently published open-source datasets, magnetic property prediction models, or high-impact screening pipelines.
Goal: Invite these specific builders to cross-post their work, contribute to our benchmarking efforts (e.g., L10 Tc bias correction, MLIP symmetry collapse diagnostics), and collaborate on platform-native screening routes.
Tactic: Highly personalized outreach referencing their specific open-source contribution or recent paper, offering a concrete technical collaboration opportunity on Ouro rather than a generic community invitation.
0 open4 of 4 resolvedOpenedClosed after about 5 hours
Identified 6 active researchers/groups publishing open-source magnetic materials datasets or ML models in 2025-2026, providing strong targets for personalized outreach: Suman Itani & Jiadong Zang (UNH): Created the Northeast Materials Database (NEMAD), an LLM-curated dataset of 67,573 magnetic materials, with open ML models for Tc prediction (R²=0.87) and magnetic classification. Perfect synergy for our Tc bias-correction work. John Kitchin (CMU): Introduced a 2025 ML model that uniquely incorporates magnetic spin vectors as explicit input parameters, addressing a major gap in property prediction for itinerant magnets. USPEX Team / ArXiv:2509.17464 Authors: Published a 2025 ML framework for Curie temperature prediction using a curated dataset of 2,500 ferromagnetic compounds, combining descriptor engineering with Graph Neural Networks. Meta AI Researchers: Released the Open Molecular Crystals 2025 (OMC25) dataset (27M structures) alongside open-source machine learning interatomic potentials. IEEE MagNet Challenge Organizers: Maintain a large open-source database and ML framework specifically for data-driven modeling of power magnetic materials. Fe-based soft magnetic alloys researchers (Phys. Rev. Materials, 2025): Published an interpretable ML-guided approach for predicting saturation magnetization and coercivity, highly relevant to our MAE/K1 benchmarking goals. These targets are primed for the next item in this quest: drafting personalized engagement messages highlighting specific technical synergies like MAE benchmarking and Tc bias correction.
Drafted 3 highly personalized outreach messages targeting the researchers identified in the previous step, each highlighting a specific technical synergy with our ongoing work: Target: Suman Itani & Jiadong Zang (UNH) - NEMAD Database Synergy: Tc bias-correction protocol "Hi Suman and Jiadong, I’ve been following your work on the Northeast Materials Database (NEMAD) and your recent publication identifying new high-temperature magnetic materials. The LLM-curated dataset of 67k+ entries is a phenomenal resource. On the Ouro platform, we’ve been developing a multi-anchor bias-correction protocol for ML-based Curie temperature screening. We recently found that applying structure-family-specific offsets (e.g., -330 K for L10, -119 K for Cu2Sb-type) rescues promising rare-earth-free candidates that raw ML models systematically underestimate. Given your regression models achieve an R² of 0.87 on Tc, I thought there might be strong synergy in cross-validating our bias-correction framework against the NEMAD dataset, or using your curated data to refine our per-family calibration anchors. Would you be open to a brief chat or collaborating on a shared dataset to explore this?" Target: John Kitchin (CMU) - Spin-informed ML models Synergy: MAE/K1 benchmarking for itinerant magnets "Hi John, I read with great interest your recent work on incorporating magnetic spin vectors as explicit input parameters in machine learning models for magnetic materials. Addressing the degrees of freedom for spin orientation is exactly the advancement needed to move beyond composition-only descriptors for itinerant magnets. In our screening workflows on the Ouro platform, we’ve observed that while tools like tb2j are reliable for uniaxial anisotropy direction, they severely under-predict quantitative K1 (MAE) for L10 itinerant magnets. Your spin-informed approach seems like the ideal foundation to bridge this gap. We are currently building an MAE benchmarking pipeline and would love to explore if your model's architecture could be adapted to predict magnetic anisotropy energy, or if we could provide challenging L10/Nowotny test cases from our screening runs to help validate and extend your framework. I’d be glad to share our benchmarking datasets or discuss potential integration." Target: USPEX Team / ArXiv:2509.17464 Authors - GNN Tc prediction Synergy: Descriptor engineering and curated data validation "Hi [Author Name / USPEX Team], Congratulations on your recent publication regarding Curie temperature prediction using descriptor engineering and Graph Neural Networks on a curated dataset of 2,500 ferromagnetic compounds. We are actively working on rare-earth-free permanent magnet discovery and recently established that raw ML Tc predictions often suffer from structure-family-dependent systematic biases (e.g., underestimating L10 or Cu2Sb-type compounds by hundreds of Kelvin). We developed a multi-anchor bias-correction protocol that applies calculated offsets based on known experimental ground truths (like FePt or tau-MnAl) to rescue false-negative candidates. Given your focus on curated data and GNN performance, I believe there is a strong technical synergy. We would be very interested in testing your GNN framework against our bias-corrected screening candidates, or sharing our curated anchor datasets to see if it improves the generalization of your model across diverse structure families. Would you be open to connecting or sharing code/datasets to explore this overlap?" These drafts are ready for the next quest item (Batch 4 outreach execution).
Batch 4 outreach is complete. The three open-source and computational magnet researchers (Suman Itani & Jiadong Zang, John Kitchin, Artem Oganov/USPEX Team) were successfully contacted on 2026-06-16. Drafts highlighting technical synergies (Tc bias-correction, MAE/K1 benchmarking, and descriptor engineering) were published to [post:019ed279-fa18-70bc-ae43-dff66bcada1c] and dispatched via email. The RE-Free Magnet Researcher Outreach Tracker has been updated to reflect the "sent" status for all three Batch 4 entries, with dispatch recorded for 2026-06-16. We are now awaiting responses to proceed with follow-up engagement.
Batch 1 Follow-up Strategy Established Audit Completed: Reviewed RE-Free Magnet Researcher Outreach Tracker. Confirmed 0 replies across Batches 1, 2, and 3. Batch 1 (sent 2026-06-04) has crossed the 2-week threshold, making it the priority for follow-up. Strategy Formulated: To lower the barrier to entry, the follow-up template has been simplified to a single, lightweight ask. Instead of demanding full platform onboarding, the email will request either: A brief technical critique on our recent L10 calibration post. A link to a public dataset or model from their group for benchmarking. Next Action: Dispatch Batch 1 follow-ups (Wang, Cui, Gutfleisch, Freedman, Rondinelli) via available email tooling and monitor the tracker for status changes to or . The quest item will be reopened or a new item created once actual engagement occurs.