If your strategic radar for where AI is producing visible output treats general AI (large language models, image generation, coding) as the dominant signal and scientific AI as a slower-moving secondary application area, the 2024–2025 publication data shows scientific AI is moving at one of the fastest research rates of any AI application category. 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%. The growth is not concentrated in one discipline. Physical sciences (roughly 33,000 publications, +27%) and life sciences (roughly 29,000, +28%) followed similar trajectories. Earth science (smallest in absolute terms at roughly 20,460) grew 23%. The cross-discipline AI publication growth is becoming a structural feature of scientific output.

Number of AI-related publications in natural sciences, 2010–25:

The growth pattern has three structural components.

The first component: growth is sustained, not bursty. The 26% one-year increase follows comparable growth in 2023–2024 (similar pace) and 2022–2023. The trajectory has been roughly linear in absolute terms for the past three years, with the absolute numbers reaching scale where 26% annual growth represents tens of thousands of new publications per year. The pattern suggests AI in science is in the rapid-adoption middle phase of an S-curve, not an early-spike or late-plateau phase.

The second component: cross-discipline breadth. The 26% aggregate growth could in principle reflect concentration in one or two disciplines. The 27–28% growth in both physical and life sciences, and the 23% growth in Earth science, show the growth is broad. AI is being adopted across the natural sciences at roughly comparable rates rather than being driven by one early-adopter discipline pulling up the aggregate.

The third component: absolute volume now exceeds the size of many entire scientific fields. 80,150 AI-related natural science publications in 2025 is approximately the size of physics, chemistry, or biology publishing in earlier eras. AI scientific research has moved from "niche application" to "comparable in scale to established disciplines" within a 5–7 year period. The strategic implications differ at this scale than they did at the earlier scale.

Three structural drivers explain the trajectory.

The first driver: foundation models for scientific applications matured. AlphaFold 2 (2021), AlphaFold 3 (2024), and successor cofolding models opened structure prediction at scale. Domain-specific foundation models in chemistry (ChemDFM), physics (GPhyT, PhysiX), Earth science (FourCastNet 3, WeatherNext 2), and biology (Evo 2, AlphaGenome) became available in 2024–2025. Each release expanded what researchers could investigate, driving research output.

The second driver: dataset infrastructure expanded substantially. The 2025 dataset releases, OpenGenome2 (9.3 trillion base pairs), Multimodal Universe (100TB astronomical data), OMol25 (100M+ DFT calculations), Tahoe-100M (single-cell sequencing), BaseData (9.8 billion genes), represent training data infrastructure that researchers can build on. Models trained on better data produce stronger results, which drives more research output building on those results.

The third driver: scientific publishing infrastructure embraced AI-driven research. Web of Science, the database used for this analysis, captures AI-tagged publications systematically. Major journals (Nature, Science, Cell, Physical Review Letters, npj Climate Atmospheric Science) regularly publish AI scientific research. Conference venues (ICML, NeurIPS, ICLR) increasingly include scientific AI tracks. The publishing pathway is well-established and accommodates substantial research output volume.

AI in natural sciences is now one of the most active research domains globally. The 80,150 publications in 2025 represents work by tens of thousands of researchers across hundreds of institutions and dozens of countries.

Three structural implications follow for organisations whose strategic positioning intersects with scientific AI.

The first implication: scientific AI now competes with general AI for research talent. Researchers who could work on either domain are increasingly choosing scientific AI because of the visible research output, sustained funding, and direct applications. The talent flow has implications for general AI research capacity. National science funding agencies, universities, and commercial AI research labs are all engaged in this competition for talent.

The second implication: pharmaceutical, materials, and other industry sectors that depend on scientific research are operating against a substantially expanded research pace. The 80,150 publications represent a research base substantially larger than the cumulative pre-2020 output. Industries that depend on translating scientific research into commercial application (pharma, materials manufacturing, agriculture, energy) face a faster-moving research landscape. Strategic capability for engaging with this research pace is operationally important.

The third implication: national research infrastructure investment shapes long-term competitive positioning. The countries and research organisations producing the highest volume and quality of scientific AI research will shape the next generation of science-dependent industries. The 26% annual growth means the cumulative effect over 2024–2030 will produce a research base substantially larger than what exists today. National investment in scientific AI research infrastructure (compute, data, talent, regulatory frameworks) has long-term economic consequences.

Through 2026–2028, AI scientific research is likely to continue compounding at 25–30% annual growth rates, so organisations setting research and innovation strategy in 2026 should treat scientific AI as a structurally large and growing domain rather than a peripheral application area: investing in the capability to engage with the accelerating pace (through internal capability, partnerships, or research-monitoring infrastructure), planning for cumulative research-base growth (the cumulative AI scientific publication base since 2020 will likely exceed 400,000 by 2028), and engaging with the geographic and institutional distribution of that research (covered separately). The cumulative effect by 2030 will produce a research base structurally larger than the cumulative output through 2024, and the pharmaceutical, materials, energy, and adjacent industries that operate on it will look structurally different by 2030 than they do today.

The signal is direct: AI in natural sciences is one of the fastest-growing research domains globally, the growth is broad across disciplines, the absolute scale is now substantial, and the trajectory shows no signs of slowing. Strategic positioning decisions made in 2026 will shape competitive outcomes for the next decade in industries that depend on scientific research output.