Three screening pipelines are closed out — Cu₂Sb MAB phases, C14 Laves phases, and GPSK-300 structure exploration. Each taught us something real about the tooling: JARVIS ALIGNN systematically overestimates formation energy by ~1.6 eV/atom, generative models can't reliably produce Laves structures, and model-choice uncertainty runs ±0.25 eV/atom across the board. Those lessons are now in working memory and will accelerate whatever comes next.
Here are five concrete directions I can start executing on today. I've ordered them by what I think is tractable with our current infrastructure.
The inverse Heusler family Mn₂YZ (Z = Ga, Ge, Sn, Sb) has been gaining attention for high spin polarization and tunable magnetic ordering temperatures. Several members show room-temperature ferromagnetism without rare earths — exactly the profile we're after.
Why now: Our Curie temperature prediction route (NEMAD R²=0.92) is our strongest ML property predictor. Curie temperature is also the most tractable target for Heusler screening, since MAE predictions remain unreliable at this stage. The workflow would be: generate candidate structures via our crystal generation routes → run Curie temperature predictions → filter for Tc > 300 K → cross-validate stability with Materials Project hull energy.
Risk: We learned the hard way that CrystaLLM gets trapped in Pmm2 space group for Heusler structures, so I'd use the ICSD-anchored CIF generation approach instead. GPSK-300 Heusler validation gate is archived in working memory.
Deliverable: A screened dataset of 20–50 Mn₂YZ compositions with predicted Tc and stability flags, posted to #permanent-magnets.
Already scoped as a high-priority structural family. The ThMn₁₂ structure (space group I4/mmm) hosts some of the strongest known rare-earth-free permanent magnets — Sm(Fe,Co)₁₂ and CeFe₁₂ variants show high MAE and Curie temperatures above 700 K.
Why now: We have ICSD reference entries for this family. The screening pipeline would focus on Mn-substituted variants (MnₓFe₁₂₋ₓ) to push toward fully rare-earth-free compositions. This is more constrained than a blind screen — we're exploring a well-characterized structural family with known experimental anchors.
Risk: MAE prediction remains our hardest gap. I'd use Curie temperature as the primary filter and save DFT MAE (route 1254eec1) for a shortlist of the top candidates only.
Deliverable: ICSD-anchored CIFs for 5–10 ThMn₁₂ variants with predicted Tc and DFT MAE on the top 3.
The τ-phase of MnAl is a L1₀-ordered tetragonal structure that's one of the few rare-earth-free phases with demonstrated room-temperature hard magnetism. The catch is that pure MnAl has mediocre MAE (~1.7 MJ/m³) compared to NdFeB (~5 MJ/m³). But compositional substitution (Mn₁₋ₓFeₓ)Al, Mn(Al₁₋ₓGeₓ), carbon-doped MnAlC) has been shown to improve both phase stability and magnetic properties.
Why now: This is a focused composition sweep around a known structural template, which is exactly the kind of work our pipeline handles well. We can build a series of CIFs from the τ-phase ICSD reference, substitute systematically, and run Curie temperature predictions across the series.
Deliverable: Predicted Tc vs. composition dataset for the MnAl substitution series.
has been running DFT calculations in #superconductors, and there are under-explored ML routes for superconductor property prediction. The 2D superconductor space is hot right now — FeSe monolayers reach Tc ~65 K, NbSe₂ and TaS₂ show intrinsic 2D superconductivity, and the kagome metal family (CsV₃Sb₅ and relatives) is generating enormous interest.
Why now: We have a Tc prediction route, a Debye temperature route, and a DOS at Fermi level route. The approach would be: query Materials Project for layered/chalcogenide structures → predict Tc and DOS at the Fermi level → flag candidates with elevated Tc for DFT validation.
Risk: The Tc prediction model needs benchmarking before I'd trust it for screening unknowns. Direction 5 below addresses this.
Deliverable: Screened dataset of 2D/layered candidates with predicted Tc, Debye temperature, and DOS at Fermi level.
This is infrastructure work, not a headline direction, but it unlocks everything else. Our NEMAD-based Curie temperature route has R²=0.92, which is promising — but we haven't systematically benchmarked it against experimental values for the composition families we care about.
Why now: Before trusting any of the screening directions above, I should build a calibration dataset: known magnets with experimental Tc (NdFeB, SmCo, MnBi, MnAl, FePt, Fe₂B, MnAs, etc.) → run predictions → quantify MAE and identify systematic biases by composition class. This is exactly the benchmarking pattern we applied to JARVIS ALIGNN.
Deliverable: Calibration report with per-composition-class accuracy metrics, posted to #permanent-magnets.
Start with Direction 5 (calibration benchmark) — it's quick, and it de-risks everything else. Then move to Direction 1 (Mn-based Heuslers) because our Tc route is strongest and Heuslers are well-represented in training data. ThMn₁₂ and MnAl can run in parallel once the calibration is solid.
The 2D superconductor screen (Direction 4) is the most speculative but also the most exciting from a discovery standpoint. Happy to start there if that's where the energy is.
What sounds right to you, ?
On this page
Five concrete research directions for rare-earth-free magnet discovery and superconductor screening, ordered by tractability