Matscholar: The search engine for materials science researchers
summary
Search engine indexes text from 5 million abstracts to extract millions of properties using a mix of supervised NLP, bespoke language models and embeddings, and serves results via a Vespa.ai backend. A chemical-aware tokenizer and entity extractor handles formulas and name variants, embeddings cluster by application (thermoelectrics, batteries) and enable similarity queries and t-SNE visualizations. Site is privately used by hundreds worldwide, data and tooling are modular on GitHub, and a published result rediscovers overlooked materials.
AI-written summary · about
chapters (9)
- 0:00Intro
- 0:59Materials Discovery Enables Tech Breakthroughs
- 5:47Named Entity Recognition Pipeline
- 6:37Using Large Language Models for NER
- 12:54Plotting the Embeddings in 2D
- 14:33The periodic table of element embeddings
- 16:28Algorithm can infer elemental properties
- 17:19Mapping the Space of Materials
- 18:42Insights about composition from embeddings
related talks
41mAccelerated Discovery to Knowledge using Machine Learning
42mThe Machine Understands Text — Learning and Search at Micro and Macro Scale
40mVespa.ai’s Personalized Search: Advanced Ranking & Tensor framework
26mStreamlining Property Search Across Multiple Platforms
32mRetrieval Augmentation and Semantic Search at Scale
01h 30mBuilding a Search Engine
31mTaking AI/ML to the next level: Text!
57mStartup Showcase