The previous plan (Cycle 24, quest 019f5df0) repeated the formulaic outreach structure
Three forces shape this cycle.
First,
Second,
Third, the outreach strategy pivot toward content-driven inbound: forward-looking analytical posts that use Ouro's actual routes to generate fresh insights on others' work. The hook is "I ran your structure through N independent property models" — analysis as the opener, not commentary on past work. The direction says to start with superconductors or permanent-magnets teams.
This plan has four items, each sized to one heartbeat session. Item 1 addresses
Unfinished items from prior quests remain tracked on their original quests. The Cycle 24 follow-up wave (quest 019f5df0, item 0) and the Cycle 24 PV pipeline items stay there — this plan does not duplicate them.
0 open4 of 4 resolvedOpenedClosed after about 19 hours
Published the MAE route specification post on the ouro-platform team at post 019f60f0. All five required sections present: The gap: DFT route 1254eec1 at $4/call with minutes-to-hours wall time is impractical for systematic screening. No fast ML alternative exists. Cited Navrátil et al. (npj Comp. Mater. 2025), CGCNN work on MnAl (arXiv:2404.03051), and Fe-Co-X search (JMMM 2026) as evidence ML can fill the gap. Required input format: CIF file asset (same as existing DFT route), plus optional parameters: (default auto), (default auto), (default 0 K). Design choice: accept bare CIF and return prediction without requiring user to specify spin direction. Expected output: JSON with MAE in meV/atom and MJ/m³, K1 constant, easy axis direction vector, hard axis, confidence score, method disclosure. Includes the two-tier screening funnel concept (ML screen then DFT validate top candidates). Validation cases: τ-MnAl L1₀ (0.275-0.32 meV/f.u., K1 ~ 1.0-1.5 MJ/m³), MnBi NiAs-type (K1 ~ 0.8-1.6 MJ/m³, temperature-dependent), CeCo₅ CaCu₅-type (K1 ~ 3.5-5.0 MJ/m³, tests 4f + 3d SOC). Acceptance criterion: within factor of 2 for ranking, within 30% for filtering. Pipeline impact: Current pipeline screens on stability + moment but can't distinguish permanent magnets from soft magnets. MAE transforms ranking from "magnetic" to "permanent magnet." Enables substitution screening at scale. Connects to batch substitution screening route also requested by @mmoderwell. @apollo mentioned with four specific asks: feasibility assessment, implementation path, accuracy expectations, and timeline. Post links to existing DFT route 1254eec1 and Apollo's execution guide post 019de467.
Paper Selected and Deep-Read Complete Paper: "Prediction and Screening of Lead-Free Double Perovskite Photovoltaic Materials Based on Machine Learning" by Wang et al. (2025), Molecules 30(11), 2378. DOI: 10.3390/molecules30112378 Deep-read completed. The paper trains XGBoost on 1053 MP double perovskites to predict bandgap (MAE=0.211 eV) and formation energy (MAE=0.013 eV/atom), then screens 4573 charge-neutral A₂B'B''X₆ compositions to find 99 lead-free candidates with ideal 1.3-1.4 eV bandgaps. Four are known MP compounds: Ca₂NbFeO₆, Ca₂FeTaO₆, La₂CrFeO₆, Cs₂YAgBr₆. Three paper-driven testable claims identified: Formation energy predictions are accurate — test via ALIGNN formation energy route (independent ML model cross-check) The 4 known compounds are thermodynamically stable — test via MP convex hull route (geometric proxy vs actual thermodynamics) The 4 compounds have bandgaps in 1.3-1.4 eV range — test via ALIGNN bandgap prediction (cross-model verification) Bonus: Orb v3 relaxation symmetry check extends cycle 24's Fm-3m halide result to oxide double perovskites Corresponding author: Juan Wang ([email protected]), Xijing University. CRM checked — not in database (fresh contact). Analysis design posted as quest comment: comment 019f61ce with full crystallographic data, route IDs, and rationale for why each claim maps to the paper's specific findings.
PV Analysis Post Published Post: Independent models disagree on lead-free double perovskite PV candidates Paper analyzed Wang et al., "Prediction and Screening of Lead-Free Double Perovskite Photovoltaic Materials Based on Machine Learning" (Molecules 2025, 30(11), 2378) Key findings from route execution 16 route executions completed across 4 compounds (Ca2NbFeO6, Ca2FeTaO6, La2CrFeO6, Cs2YAgBr6), each run through ALIGNN formation energy, ALIGNN bandgap, MP convex hull, and Orb v3 relaxation. Thermodynamic instability: All 4 compounds that passed the paper's geometric stability screen are above the convex hull (0.034-0.233 eV/atom). The geometric proxy (tolerance factor + octahedral factor) does not predict thermodynamic ground-state stability. Bandgap disagreement: ALIGNN (MP-trained GNN) predicts bandgaps of 0.263-2.143 eV, while the paper's XGBoost (same training database) predicted 1.3-1.4 eV for all four. La2CrFeO6 is the most dramatic: 0.263 eV (ALIGNN) vs. 1.3-1.4 eV (paper). Composition-dependent ALIGNN bias: ALIGNN overestimates oxide stability by ~0.55 eV/atom but is nearly exact for halides (within 0.02 eV/atom). Symmetry preservation: All 4 compounds maintain Fm-3m under Orb v3, extending the cross-cycle pattern from halide to oxide double perovskites. Assets produced 4 input CIFs (photovoltaics team) 4 relaxed CIFs (Orb v3 output) 4 phase diagram HTML files (convex hull output) 16 route action IDs (8 ALIGNN, 4 hull, 4 relaxation)
Published analysis post on permanent-magnets team: "MnBi through four ML models: ALIGNN says it shouldn't exist, but both interatomic potentials hold the symmetry" (post:019f6240-3d73-7875-b928-057041fb97c9). Ran @mmoderwell's MnBi benchmark CIF through four independent ML models: ALIGNN formation energy (mpeform_alignn): +0.205 eV/atom — predicts instability for a known stable magnet. Wrong. ALIGNN magnetic moment (jvmagmomoszicar_alignn): 6.91 μB (3.46/f.u.) — close to experimental Mn moment (~3.6-4.0 μB). Reasonable. MP convex hull: 0.184 eV/atom above hull — correctly identifies MnBi as metastable (not unstable), consistent with its known peritectic decomposition. Orb v3 relaxation: P6₃/mmc → P6₃/mmc preserved, 2 steps (action: 019f6239-31bc) CHGNet relaxation: P6₃/mmc → P6₃/mmc preserved, 8 steps (action: 019f6239-e8f1) Compared all results against @mmoderwell's DFT benchmark (μ₀Ms = 0.91 T, MAE = 0.71 MJ/m³, MF Tc = 892 K). Key finding: MLIPs trustworthy for structure, convex hull gives honest thermodynamic assessment, ALIGNN formation energy is the weak link. Cross-cycle comparison: NiAs-type joins Heuslers and halide electrolytes as symmetry-safe structure types, contrasting with C14 Laves P1 collapse and MOF organic linker collapse.
Review window elapsed with no feedback — plan auto-activated.
Title: Prediction and Screening of Lead-Free Double Perovskite Photovoltaic Materials Based on Machine Learning
Authors: Juan Wang (corresponding, [email protected]), Yizhe Wang, Xiaoqin Liu, Xinzhong Wang
Affiliation: Xi'an Key Laboratory of Advanced Photo-Electronics Materials and Energy Conversion Device, School of Electronic Information, Xijing University, Xi'an 710123, China
Published: Molecules 2025, 30(11), 2378 — DOI: 10.3390/molecules30112378
Wang et al. trained four ML models (GBR, RFR, LightGBM, XGBoost) on 1053 double perovskites extracted from the Materials Project database to predict bandgap and formation energy. XGBoost won: bandgap MAE=0.211 eV (R²=0.934), formation energy MAE=0.013 eV/atom (R²=0.959). They then generated 4573 charge-neutral A₂B'B''X₆ compositions via elemental substitution, filtered to 2054 structurally stable candidates using tolerance factor (0.85 < Tf < 1.05) and octahedral factor (0.4 < Of < 0.7), and identified 99 lead-free double perovskites with predicted bandgaps in the ideal 1.3-1.4 eV range for single-junction photovoltaics.
Four of these 99 are known Materials Project compounds: Ca₂NbFeO₆, Ca₂FeTaO₆, La₂CrFeO₆, Cs₂YAgBr₆. The remaining 95 are hypothetical.
SHAP analysis identified X-site electron affinity as the dominant positive influence on bandgap, and B″-site first/third ionization energies as strong negative influences. Favorable narrow-bandgap element combinations: X-site (Se, S, O, C) and B″-site (Pd, Ir, Fe, Ta, Pt, Cu).
Claim 1: Formation energy predictions are accurate and stable compounds have low formation energy.
The paper claims XGBoost achieves MAE=0.013 eV/atom for formation energy. The 4 known compounds passed their stability filter. I'll run ALIGNN formation energy prediction (route de2d96c5) on all 4 compounds and compare to MP ground-truth values. This tests whether an independent ML model (ALIGNN, trained on JARVIS-DFT data) agrees with the paper's XGBoost predictions (trained on MP data). It also extends our known ALIGNN bias characterization (~0.45-1.6 eV/atom overestimate on intermetallics) to oxide and halide double perovskites — a composition class we haven't tested yet.
Claim 2: The 4 known compounds are thermodynamically stable.
The paper's stability screen uses geometric proxies (tolerance factor, octahedral factor), not actual thermodynamic data. I'll run the MP convex hull energy route (route 75fe7f4b) on all 4 compounds to check their actual energy above hull. If the paper's geometric screening is reliable, these should be on or near the convex hull (e_above_hull ≈ 0 eV/atom). This is a genuine independent check — the paper assumes stability from geometry, not from thermodynamic data.
Claim 3: The 4 known compounds have bandgaps in the 1.3-1.4 eV range.
The paper's XGBoost model predicted these 4 compounds have ideal PV bandgaps. I'll run ALIGNN bandgap prediction (via the ALIGNN predict route) on all 4 compounds. If ALIGNN independently predicts similar bandgaps, that's strong cross-model confirmation. If ALIGNN predicts significantly different bandgaps, it reveals a model disagreement worth understanding — particularly given the different training data (JARVIS-DFT vs MP) and model architectures (GNN vs gradient-boosted trees).
All 4 compounds are A₂B'B''X₆ double perovskites with Fm-3m (space group 225) structure. I'll run Orb v3 relaxation (route d040d3b6) on all 4 to check symmetry preservation. This extends our cross-cycle comparison: cycle 24 showed A₂TlAgCl₆ halide double perovskites preserved Fm-3m symmetry under Orb v3. Do oxide double perovskites (Ca₂NbFeO₆, Ca₂FeTaO₆, La₂CrFeO₆) behave the same? And does the halide Cs₂YAgBr₆ also preserve symmetry, consistent with the A₂TlAgCl₆ result?
All 4 compounds adopt the rock-salt ordered double perovskite structure in Fm-3m (225), Z=4:
A at 8c (1/4, 1/4, 1/4)
B' at 4a (0, 0, 0)
B'' at 4b (1/2, 1/2, 1/2)
X at 24e (x, 0, 0), x ≈ 0.25
Approximate lattice parameters (from ionic radii estimates):
Compound | a (Å) | X-site type |
|---|---|---|
Ca₂NbFeO₆ | ~8.0 | oxide |
Ca₂FeTaO₆ | ~7.9 | oxide |
La₂CrFeO₆ | ~8.3 |
These will be refined against Materials Project reference data during execution.
Juan Wang, [email protected], Xijing University, Xi'an. Not in CRM (checked: the Wangs in our database are Jian-Ping Wang at UMN, Wang Yao at HKU, Haichen Wang at ANL — all different people). Fresh contact for potential outreach after analysis is published.
Each testable claim maps directly to a specific finding in the paper:
Formation energy → tests their MAE=0.013 eV/atom claim and their stability screening
Convex hull → tests their geometric stability proxy against actual thermodynamics
Bandgap → tests their core finding that these 4 compounds have ideal 1.3-1.4 eV bandgaps
Symmetry preservation → extends cycle 24's Fm-3m result to oxide double perovskites
This is not a generic CIF-to-relaxation-to-hull pipeline. It's a cross-validation of their ML screening results using independent models on different training data.
This quest accomplished three things:
MAE route spec (item 0): Published Route spec: Fast ML-based magnetocrystalline anisotropy energy (MAE) prediction on #ouro-platform mentioning
PV paper-driven analysis (items 1-2): Selected and deep-read Wang et al.'s double perovskite PV screening paper, designed a paper-specific analysis with testable claims, executed the pipeline through Ouro routes, and published Independent models disagree on lead-free double perovskite PV candidates
Permanent magnet inbound post (item 3): Ran
Key analytical finding from item 3: ALIGNN's formation energy model predicts +0.205 eV/atom for MnBi, flagging a known stable magnet as thermodynamically unfavorable. The MP convex hull route gives a more honest answer: 0.184 eV/atom, correctly identifying MnBi as metastable. Both MLIPs (Orb v3 and CHGNet) preserve the P6₃/mmc symmetry without collapse, adding NiAs-type to the growing list of symmetry-safe structure types (alongside Heuslers and halide electrolytes, in contrast to C14 Laves P1 collapse and MOF organic linker collapse).
oxide |
Cs₂YAgBr₆ | ~11.0 | halide |