In the Web of Science database, AI-related publications in the natural sciences reached approximately 80,150 in 2025, up from 63,547 in 2024 — a one-year increase of roughly 26%. Physical sciences and life sciences followed similar trajectories, reaching about 33,000 and 29,000 publications respectively, each growing 27%–28% year over year, while Earth science — the smallest category at about 20,460 — grew roughly 23%.

AI is becoming routine across disciplines

• As a share of total output, AI-related work remains a single-digit fraction of each field but is climbing fast. By 2025, Earth science had the highest AI penetration at 8.8%, followed by natural sciences overall at 6.8%, life sciences at 6.5%, and physical sciences at 5.8%. In 2010, all four categories sat below 1%.
• The clearest breakthroughs cluster in domains with strong existing data infrastructure — structural biology, physics, chemistry, and materials science — rather than in fields built on the most sophisticated mathematical or physics-based models.
• These developments do not automatically translate into scientific progress. Experimental validation remains expensive and slow: AI can propose novel candidate molecules at scale, but clinical trials to determine whether they work remain a costly, multiyear process. The gap between what AI can propose and what scientists can feasibly test recurs across every domain.

Physics, astronomy, chemistry, materials — and a small-model surprise in biology

• In molecular biology, smaller models are outperforming larger ones. MSAPairformer, a 111-million-parameter protein language model, surpassed previous leading methods on the ProteinGym benchmark; and GPN-Star, a 200-million-parameter genomics model, outperformed Evo 2, a model with 40 billion parameters.
• Virtual-cell models emerged as a new 2025 frontier: Evo 2 (Arc Institute), STATE, and DeepMind's AlphaGenome aim to predict cellular responses to drugs and genetic perturbations without wet-lab experiments — though current systems still require experimental validation. Evo 2 trained on OpenGenome2, a corpus of 9.3 trillion DNA base pairs from across all domains of life.
• Astronomy released its first foundation model, first visualization benchmark, and a 100TB training dataset in 2025. AION-1, trained on over 200 million celestial objects from five major surveys, is the field's first foundation model; AstroVisBench introduced the first benchmark for LLM scientific computing and visualization.
• An AI system ran a full weather-forecasting pipeline end-to-end for the first time in 2025: Aardvark Weather replaced the traditional numerical pipeline with a single ML system, and multiple AI weather models reached operational deployment. FourCastNet 3 generates a 60-day global forecast in under 4 minutes — 8 to 60 times faster than prior approaches.

Strong on subtasks, shaky on full research

Frontier models outperform human chemists on average but cannot reproduce published research. On ChemBench (2,700+ chemistry questions), the best models beat human-expert averages while still stumbling on basic tasks. On ReplicationBench, frontier models score below 20% on paper-scale replication in astrophysics. On UnivEarth, LLM agents answer Earth-observation questions with just 33% accuracy and their code fails 58% of the time, while on BixBench, frontier models reach only ~17% on real-world bioinformatics analysis.