SciAtlas
Inspiration
Have you ever tried reading a research abstract, only to give up immediately because it felt completely incomprehensible?
Or perhaps you’ve asked an AI chatbot for help researching an unfamiliar topic, only to receive an answer that was confusing, outdated, or flat-out wrong?
Human knowledge is dynamic, interconnected, and multi-dimensional. Yet for thirty years, the way we navigate the internet has remained flat and linear: a search bar, a wall of blue links, endless browser tabs, and information overload.
Whether you’re a student, novice researcher, or curious hobbyist, let's face it: doing research and delving into technical topics on the internet can be a frustrating and limiting process. SciAtlas is a web application that redesigns what it means to navigate information on the internet by turning the current research landscape into a searchable 3D galaxy that highlights not just answers, but scientific consensus, counterarguments, and gaps in knowledge.
What it does
When a user searches for a question or domain, SciAtlas retrieves roughly 100 relevant research papers and generates an interactive, 3D galaxy-inspired map to explore.
- Nodes & Clusters: Each paper is rendered as a "star," with closely related studies clustering into bright constellations. The largest cluster represents the scientific consensus answering the user's query.
- Visual Cues: Node size and brightness reflect citation popularity—larger, brighter nodes represent heavily cited papers, while dimmer nodes indicate less popular works.
- Gaps & Contradictions: Crucially, SciAtlas highlights contradicting studies alongside gaps in current literature where further research is needed.
- Adaptive Summaries: Everything is labeled dynamically across three user-adjustable levels of understanding: Curious, Student, and Expert.
How I built it
- Data Retrieval: Integrated the OpenAlex API to access its index of over 200 million research papers.
- Embeddings: Converted relevant metadata (title, author, abstract, citations) into 768-dimensional vector embeddings using the SPECTER 2 proximity model on Hugging Face.
- Dimensionality Reduction & Layout: Applied UMAP to reduce high-dimensional vectors for spatial representation.
- LLM Integration: Utilized the Gemini API to generate customized text summaries for clusters and individual papers tailored to the selected expertise level.
- UI: Rendered the interactive 3D visualization and interface in Python using Streamlit and
3d-force-graph.
Challenges I ran into
The primary technical challenge was ensuring that visual gaps in the 3D map accurately corresponded to actual gaps in research literature. Because UMAP reduces high-dimensional data into 3D coordinates, cluster distances and gap sizes in the visual rendering can become somewhat arbitrary.
To solve this, literature gaps and clusters could not be calculated purely from 3D visual coordinates; they had to be computed directly within the original high-dimensional vector space using K-Nearest Neighbors (KNN) and K-Means clustering before mapping them visually.
Accomplishments that I’m proud of
I am particularly proud of the user interface and how much information is seamlessly packaged into a single view. A single search query unfolds a wealth of insight through spatial clusters, brightness, color, gaps, and adaptive text summaries. Compare this to the standard approach of endless typing and scrolling that follow the usual methods of searching on the internet. SciAtlas is a breath of fresh air that offers a better method for exploring scientific knowledge.
What I learned
Through building SciAtlas, I deepened my understanding of high-dimensional vector spaces and the mathematical principles behind algorithms like K-Nearest Neighbors and K-Means. I also gained practical experience integrating Large Language Models into an application workflow using the Gemini API.
What's next for SciAtlas
Looking forward, I plan to make SciAtlas even more immersive by introducing automated recommendations for adjacent research domains ("galaxies") to explore. Additionally, I aim to support multi-view side-by-side visualizations and enable inter- "galaxy" navigation, allowing users to jump fluidly through one knowledge network and into another.
Log in or sign up for Devpost to join the conversation.