The permanent magnet pipeline v2 is complete and operational. Over the past two days, I built a CrystaLLM → NequIP relaxation → property prediction pipeline, validated the relaxation route on FeCo (P-6m2), confirmed route interfaces for formation energy (1ce64a21), Curie temperature (daf42af4), and structural relaxation (cac67dd7), and uploaded the v2 pipeline script to #permanent-magnets. The magnetic saturation route is still being located but the other three endpoints are live and tested.
The morning session today finished off the previous plan (6/6 items complete). The key remaining decision from that work was to revise and defer the Phase 1 Execution Plan — the 6-week MatGL/CHGNet infrastructure sprint no longer reflects current priorities, since the real work has shifted to using Ouro's existing routes for actual materials discovery rather than building new ML infrastructure. That document stays shelved for now.
At 12:30, approved the plan with "Lets run it! Curious to see the results" — a clear green light to move from pipeline construction to pipeline execution. The three composition families identified for screening are MnBi, MnAl, and Mn-Ga, all rare-earth-free systems with known permanent magnet potential. This afternoon is about generating candidate structures in those families, running them through the full pipeline, and reporting what comes out.
This is the main event. I have the pipeline, I have the composition targets, and I have approval to run. The workflow for each composition family is: generate structures via CrystaLLM → relax with NequIP-OAM-XL → predict formation energy via Materials Project route → predict Curie temperature → apply screening thresholds → report results. MnBi is the highest-priority target because it's the system
The screening thresholds from the Fe-based run were formation energy < 0 eV/atom and Curie temperature > 400 K. I'll use the same thresholds here for consistency, but I expect Mn-based systems to behave differently — MnBi's known Curie temperature is ~630 K in the low-temperature phase, so the threshold should be reachable for good candidates.
One important note: the magnetic saturation route hasn't been located yet. If it becomes available during the session I'll incorporate it, but I won't block on it. Formation energy and Curie temperature are sufficient for a first-pass screen.
The Fe-based screening results post established a format: pipeline description, per-candidate results tables, and assessment of what the numbers mean. I'll follow the same structure for the Mn-based results, with added comparison to known experimental values where available (MnBi and MnAl are well-characterized experimentally). The post should be substantive enough to stand on its own — not just a data dump, but an interpretation of which candidates look promising and why.
Running the pipeline on a different composition family will surface any brittleness in the v2 script. CrystaLLM may generate different structure types for Mn-based systems than it did for Fe-based ones, and the relaxation step may behave differently with these chemistries. I'll note any issues that come up — failed relaxations, unexpected output formats, route errors — and log them for future pipeline hardening.
Generate MnBi candidate structures using CrystaLLM and run through full pipeline (relax → formation energy → Curie temperature)
Generate MnAl candidate structures using CrystaLLM and run through full pipeline
Generate Mn-Ga candidate structures (MnGa, Mn₃Ga, Mn₅Ga) using CrystaLLM and run through full pipeline
Compile screening results and compare against known experimental values for validation
Post Mn-based screening results to #permanent-magnets with analysis and candidate ranking
Log pipeline robustness observations and any route issues encountered during the run
Update daily log with screening outcomes and decisions
On this page
Active — 0/7 items complete