It's been a productive day. The Mn-based rare-earth-free magnet screening campaign (Plan 2bbe4743) is fully complete — all seven items done. MnBi, MnAl, and Mn-Ga compositions were each run through the CrystaLLM → NequIP-OAM-XL relaxation → Materials Project formation energy → Curie temperature pipeline, and results were posted to #permanent-magnets with experimental validation comparisons. MnBi in the P-6m2 structure came out as the top candidate with a formation energy of 0.030 eV/atom and a predicted Curie temperature of ~1115 K. has been engaged throughout the day and approved the screening run enthusiastically.
The most important finding from today's work isn't a candidate — it's a calibration problem. The Curie temperature route (daf42af4) shows large systematic discrepancies against known experimental values: MnBi predicted at 1115 K versus the experimental range of 540–630 K, MnAl at 462 K versus ~650 K experimental, and Mn₅Ga at 268 K versus 470–770 K experimental. The errors aren't even consistent in direction — MnBi overshoots by nearly 2×, while MnAl undershoots by ~30%. This means the Curie temperature predictions are useful for rough ordering but not for absolute thresholds. Any screening that gates on a hard T_C cutoff is unreliable right now.
The pipeline v2 script is uploaded and working, all route interfaces are confirmed, and the three initial composition families have been screened. The earlier plan items around wiring the pipeline, validating on Fe-based systems, and running the Mn campaign are all behind us. What remains is to figure out what to do with these results and where to point the pipeline next.
This is the highest-value next step. Before expanding the screening to more compositions, I need to understand the T_C prediction errors better. The plan is to assemble a calibration dataset — known permanent magnets with well-characterized experimental Curie temperatures — and run each through the prediction route. With enough data points I can fit a simple correction (linear regression, or even just a per-chemistry offset) that makes the predictions usable for absolute comparisons. The Fe-based screening from yesterday provides additional data points (Fe₂O₃, FeCo, Fe₃N all have known experimental T_C values). Combining those with today's Mn results gives a starting dataset of at least 6 compositions.
This is worth doing before launching another big screening campaign, because without calibration the results are hard to interpret. A 20-minute analysis that produces a correction factor is more valuable than screening 10 more compositions with uncalibrated predictions.
Three promising directions were identified in the Mn results post: Mn-Fe-X Laves phases, Mn₂YZ Heusler variants, and extended binary Mn-X systems. Of these, Heusler alloys are particularly interesting because their crystal structure is well-defined (L2₁ or C1b) and CrystaLLM should generate them cleanly. They're also a rich design space for tuning magnetic properties through composition. If time permits after the calibration work, I'll start generating a few Heusler candidates to test whether the pipeline handles the more complex unit cells.
The pipeline has been through two composition families now (Fe-based and Mn-based) and I've accumulated a set of practical observations about what works and what breaks. The route intermittency issues from this afternoon, the CrystaLLM fallback to OMatG for certain compositions, and the Curie temperature calibration gap are all worth documenting in a consolidated "lessons learned" note. This isn't urgent but it's the kind of thing that's easy to do while it's fresh and painful to reconstruct later.
Assemble calibration dataset: collect known experimental Curie temperatures for all compositions screened so far (Fe₂O₃, FeCo, Fe₃N, MnBi, MnAl, Mn₅Ga) and compare against pipeline predictions
Fit a simple correction model for the T_C route and document the calibration approach
Post calibration analysis to #permanent-magnets or #materials-science with methodology and corrected predictions
Generate 2–3 Mn₂YZ Heusler candidates via CrystaLLM and run through the pipeline as a test of the next screening direction
Write a brief pipeline lessons-learned note covering route reliability, CrystaLLM vs OMatG fallback behavior, and prediction accuracy observations
Update daily log with calibration results and any new screening outcomes
On this page
Pending review — will auto-activate at 20:01