What is science actually working on? I built three open-source trackers to find out.

We all feel the firehose: thousands of papers a month, every field moving at once, and no honest way to tell a genuine wave from your own reading bias. So I built a small family of tools that answer one question with data instead of vibes:

Which research topics are rising, which are fading, and what are the papers behind the trend?

Three sibling projects, one shared engine, three fields:

All three are free, open-source (MIT), and live right now — static sites, no login, no tracking.

What each one focuses on

🤖 AI-Trends — what the field is presenting. Conferences are where AI commits to the record, so this one ingests every accepted paper from the big five venues and labels it by research topic. You get top / emerging / fading topics year over year, plus a Rising Stars view that sorts papers by Semantic Scholar citations (title-verified to avoid wrong matches) and links each to arXiv / Google Scholar / the source PDF.

🧬 Bio-Trends — what biology is publishing right now. Journals move continuously, so this one polls RSS feeds from the flagship biology families and accumulates them into a deduplicated monthly store. Trends are viewable at Year / Quarter / Month granularity, and a 33-topic taxonomy derived from the literature itself (scispaCy NER over a ~59k-title corpus) keeps the labels honest.

🔬 Mater-Trends — what materials science is publishing right now. Same RSS engine as Bio-Trends, pointed at the broadest journal set of the three — seven publisher families across batteries, photovoltaics, 2D materials, catalysis, quantum materials, and more. Hand-seeded 33-topic taxonomy that expands from the corpus over time.

The two journal trackers (Bio + Mater) add a nice touch: rising-citation tracking. Each paper is snapshotted when first seen, then re-checked monthly for three months — so the gain in citations becomes an early “this one’s taking off” signal, not just a static count.

How they’re built (the part I’m actually proud of)

Every project runs the same deterministic pipeline. Swap the source and the topic list, and the same engine works for a new field:

sources ──▶ ingest ──▶ assign topics ──▶ compute trends ──▶ static site
(RSS feeds /  (accumulate +   (substring match    (top/emerging/    (GitHub Pages,
 conference   dedup by DOI     vs a curated        fading per         lazy-loaded
 paper dumps) → monthly CSV)   taxonomy)           group & period)    paper shards)

The design principles, in plain terms:

1. Config is the source of truth. Adding a journal, feed, or topic is a config edit, not a code change. The journal registry and taxonomy are just JSON.

2. Everything reproducible except one human step. Ingest, topic assignment, trend math, and site export are pure, deterministic Python — run them twice, get the same answer. The only non-deterministic step is topic curation (deciding which keyword belongs to which topic), and even that is assisted: spaCy/scispaCy NER surfaces candidate keywords from the real corpus, and a Claude Code skill (/curate-topics) decides each one — keep / new topic / noise — so the taxonomy grows from the literature and stays comparable across years.

3. Data identity comes from file location. A paper’s journal is its folder; its month is the filename. Dedup falls back DOI → link → title. Idempotent: re-running with nothing new just re-derives the same outputs.

4. The site is static and fast. The pipeline exports JSON shards (one per journal-year) and the page lazy-loads only the years you actually open, so a multi-year corpus stays browsable without a heavy initial load. It deploys straight from /docs on GitHub Pages — no server to run, nothing to pay for.

5. It keeps itself current. A monthly GitHub Actions cron polls the sources, re-runs the pipeline, and opens a PR; merging republishes the site. Journals even declare a publish cadence so feeds are polled no more often than they actually update.

Why I think this is useful

• For researchers & students — see where a field is heading before it’s obvious, and jump straight to the papers behind any trend.
• For anyone explaining a field — defensible, sourced “here’s what’s hot” charts instead of anecdotes.
• For builders — the engine is field-agnostic. Point it at a new set of RSS feeds + a topic list and you have a trend tracker for your domain in an afternoon.

Three fields are live today. The architecture is built to add a fourth.

Take a look — they’re all free and open:

• 🤖 AI → https://charlesxu90.github.io/AI-trends/
• 🧬 Bio → https://charlesxu90.github.io/Bio-trends/
• 🔬 Materials → https://charlesxu90.github.io/Mater-trends/

Stars, issues, and “please track my field next” requests all welcome. 🌟