Interactive map of 455 million academic papers — search, explore, and compare authors across the entire research landscape.
Academic Atlas turns the entire OpenAlex corpus (455M papers across 26 fields) into an interactive, semantically-zoomable map. Explore how research has evolved from 1763 to 2025, compare authors' intellectual footprints, and discover cross-disciplinary connections.
Traditional academic search returns ranked lists. But research is structured as a landscape — with continents (fields), islands (subfields), and bridges between them (interdisciplinary work). Academic Atlas renders this landscape as a map you can navigate.
Use cases:
- Researchers — find adjacent fields to your work, spot research gaps, identify potential collaborators
- PhD students — place your dissertation in the broader scholarly context
- Institutions — benchmark faculty portfolios against the global research landscape
- Recruiters / job seekers — visualize an author's research trajectory at a glance
A pre-built landscape of ~50K high-impact papers (≥500 citations) sampled via square-root proportional stratification across 26 OpenAlex fields, producing 157 topic clusters.
- Semantic zoom — cluster labels appear progressively as you zoom in (field → subfield → topic)
- GPU-rendered via deck.gl — smooth interaction with 50K+ points
- CT-scan time slider — drag through years to see research evolve. Uses deck.gl's
DataFilterExtensionfor zero-lag GPU-based filtering - Cumulative mode shows all papers up to year Y; slice mode shows only papers from year Y
Search any author → disambiguate from candidates (with institution, citation count, top paper) → confirm → build the full landscape.
2-layer ego network with automatic inner-circle detection using a composite score: frequency × recency × Jaccard similarity. Gold-bordered nodes are the focal author's closest collaborators. Hover for paper details.
KDE-based territorial map showing how an author's research spans different intellectual territories. Labels describe each territory using OpenAlex topic hierarchy (Strategy and Management, Information Systems, Organizational Behavior, etc.).
Z-axis = time. See how a scholar's research focus shifted over decades. Each dot is a paper; colors are research domains; lines connect papers in the same research thread chronologically.
Select up to 5 authors → project all their papers into a shared embedding space. Each author gets a unique marker shape (circle, diamond, square, triangle, cross) and border color. Co-authored papers appear as gold stars — instantly see where scholars overlap in research space.
Example above: Kalle Lyytinen vs Youngjin Yoo (262 papers). The gold stars reveal their co-authorship clusters across Software Engineering, Information Systems Theories, and Digital Platforms.
Type a research idea in natural language. Claude (with tool use) queries the 251M-paper SQLite FTS5 index, returns relevant papers, and you can visualize them as a focused map.
Powered by:
- Claude Sonnet 4.6 with tool use for agentic search
- SQLite FTS5 full-text index with BM25 ranking
- sentence-transformers (MiniLM-L6-v2) for 384-dim embeddings
- BERTopic / KMeans + UMAP for clustering and 2D projection
Digital Innovation scholar, 109 co-authors, 631 papers.
| Co-author Network | Research Territory | 3D Trajectory |
|---|---|---|
 |
 |
 |
The trajectory view reveals how Yoo's work expanded from early IT strategy research (~2000s) into Digital Platforms and Sustainability (~2020s), with a visible shift in research domain centroids over time.
OpenAlex Snapshot (583 GB, 455M papers)
↓ extract_to_parquet.py [20 hours]
Parquet Lakehouse (193 GB, year-partitioned)
↓ build_derived.py [~24 hours total]
├── analytics.duckdb (453 GB) — analytics queries
├── search.db (336 GB) — SQLite FTS5, 251M articles
├── authors.parquet (2.8 GB) — 109M deduplicated authors
└── worldmap_clustered.csv — pre-built world map
↓ app.py
Dash Web App (deck.gl + Plotly + Claude API)
Why Lakehouse Lite?
- Parquet = source of truth (cheap, columnar, reprocessable)
- DuckDB = analytics (group-by, aggregation)
- SQLite FTS5 = full-text search (BM25, fast)
- Every derived DB is rebuildable from Parquet — if schema changes, just rerun
| Layer | Tool |
|---|---|
| Data source | OpenAlex Snapshot (S3) |
| Storage format | PyArrow Parquet |
| Analytics | DuckDB |
| Search | SQLite FTS5 |
| Embeddings | sentence-transformers (all-MiniLM-L6-v2) |
| Clustering | BERTopic, KMeans, HDBSCAN |
| Dim. reduction | UMAP |
| Map rendering | deck.gl (via datamapplot) |
| Charts | Plotly |
| Web framework | Dash + Dash Bootstrap Components |
| LLM | Anthropic Claude API (tool use) |
| Remote access | Tailscale |
- Python 3.12+
- ~800 GB free disk (for full OpenAlex extraction)
- 32 GB RAM recommended (16 GB works with reduced sample sizes)
- Anthropic API key (optional, for the AI Research Assistant)
# 1. Download OpenAlex snapshot (~583 GB, S3)
python download_openalex.py
# 2. Extract to Parquet (~20 hours)
python extract_to_parquet.py
# 3. Build derived databases (~24 hours total)
python build_derived.py all
# 4. Build the world map (~25 minutes)
python build_derived.py worldmap
# 5. Launch the app
python app.py # → http://localhost:8050# App binds to 0.0.0.0 — any device on your tailnet can access
# Find your Tailscale IP:
tailscale ip
# Then visit: http://<your-tailscale-ip>:8050-
Square-root proportional sampling — Medicine (36M papers) and Dentistry (750K) both get represented fairly. Equal allocation over-represents small fields; pure proportional makes them invisible. Square-root is the academic standard for stratified science-mapping sampling.
-
OpenAlex topic labels > TF-IDF — TF-IDF produces label soup like "blockchain, blockchain technology, blockchainbased, technology, blockchains". OpenAlex gives "Blockchain Technology and Applications" — human-readable and semantically coherent.
-
Deduplicated cluster labels — Two clusters can both be dominantly "Technology Adoption" papers while occupying different positions in embedding space. We deduplicate labels (largest cluster gets the clean label, others fall back to TF-IDF) so each region has a unique identifier.
-
GPU-based time filtering — Re-rendering 50K points on every slider tick lags.
DataFilterExtensionchanges a single GPU uniform; the data stays loaded. Smooth even on a laptop. -
Author disambiguation before embedding — "Jun Xiang" returns 21 different people at different institutions. Show candidates with institution + top paper; require explicit Confirm before running the expensive ego-network build.
v0.5 (2026-04) — Current
- Full Lakehouse Lite (455M papers)
- World Map with sqrt-proportional sampling
- CT scan time slider (deck.gl DataFilterExtension)
- Author disambiguation + confirm flow
- Multi-author comparison with cluster labels
- Remote access via Tailscale
Roadmap (see BACKLOG.md)
- Research gap detection (click empty region → LLM analyzes potential cross-disciplinary opportunities)
- Semantic search (embedding similarity, not just keyword)
- Citation network visualization
- Docker deployment
MIT License — see LICENSE.
Built on open data (OpenAlex) and open tooling.
@jscmp4 — PhD student researching online communities, computational social science, and AI agents.