AI-accelerated mathematical discovery

2 posts · Sep 2025 – May 2026

What this trail covers

AI-accelerated mathematical discovery continues to advance, with systems like AlphaProof and Gemini's Deep Think mode achieving medal-standard performance in mathematical olympiads. These systems leverage techniques pioneered by AlphaGo, such as reinforcement learning and search algorithms, to prove formal mathematical statements and solve complex, open-ended challenges. The development of AI co-scientists further enhances this capability by enabling agents to debate scientific ideas and identify new mathematical problems. This initiative expands this by partnering with leading research institutions to build infrastructure and tools for AI-driven mathematical research.

All DeepMind trails

May 2026Gemini for Science: AI experiments and tools for a new era of discoveryThis post introduces Gemini for Science, a suite of AI tools and experiments to accelerate scientific discovery. It details three primary prototypes: Hypothesis Generation (using Co-Scientist for multi-agent idea tournaments), Computational Discovery (using AlphaEvolve and ERA for generating and scoring code variations), and Literature Insights (using NotebookLM for literature analysis and artifact creation). It also highlights the Science Skills bundle for life science databases and enterprise solutions via Google Cloud. The post mentions the publication of ERA and Co-Scientist research papers in Nature and the development of agentic peer review tools.
May 2026Co-Scientist: A multi-agent AI partner to accelerate researchIntroduces Co-Scientist, a multi-agent AI system built with Gemini that iteratively generates, debates, and evolves novel hypotheses for complex scientific problems. It features specialized agents for idea generation, debate, and evolution, orchestrated by a supervisor agent. The system uses a 'tournament of ideas' approach, drawing inspiration from AlphaGo and AlphaStar, to verify, refine, and rank hypotheses by cross-checking against scientific literature and data, integrating tools like web search, ChEMBL, UniProt, and AlphaFold.
May 2026Co-ScientistThis post details a specific application of Co-Scientist in identifying repurposed drugs for liver fibrosis. It showcases how Co-Scientist's ability to connect disparate findings from scientific literature led to the identification of two effective drug candidates, one of which was a cancer drug (vorinostat) with limited prior association with fibrosis. The experiments conducted by Gary Peltz's team at Stanford validated Co-Scientist's suggestions, demonstrating its utility in accelerating drug discovery by suggesting drugs that reshape gene activity.
May 2026Co-ScientistThis post details the application of Co-Scientist, an AI system, in biomedical research for metabolic dysfunction-associated steatohepatitis (MASH). It describes how Co-Scientist synthesized evidence from liver biology and pharmacology to identify potential drug pairings and generated a hypothesis about the NLRP3 inflammasome's role in MASH, which was later experimentally verified. This demonstrates the system's capability in hypothesis generation and literature synthesis for complex biological problems.
May 2026Co-ScientistIntroduces Co-Scientist, an AI system used by Calico Life Sciences to generate a novel hypothesis about ISR regulation by metabolism. The system also assisted in refining experimental design and incorporating new findings, leading to new discoveries in the biology of aging.
Advertisement
May 2026Co-ScientistThis post introduces Co-Scientist, an AI agent that assists biologists Omar Abudayyeh and Jonathan Gootenberg in aging research. Co-Scientist generates novel genetic factors to test by scanning scientific literature and validates hypotheses through lab tests. It also significantly reduces the time required to analyze experimental screening data, slashing analysis from six months to a few days by integrating screening results with the scientific literature.
May 2026Co-ScientistThis post details how Co-Scientist was used by Professor Clare Bryant to accelerate the identification of specific amino acids for experimental testing in sepsis and flu research. The system generated and ranked hypotheses, prioritized a previously unconsidered protein, and refined hypotheses down to specific amino acid mutations. This process, which would typically take 2-3 years, is on track to be completed in six months.
Mar 2026AlphaGo at 10: How AI Innovation Is Paving the Path to AGIThis post highlights the evolution of AI in mathematical discovery, tracing its lineage from AlphaGo's techniques to current systems like AlphaProof and Gemini's Deep Think mode. It details how these systems achieve medal-standard performance in mathematical olympiads by combining language models with reinforcement learning and search algorithms. It also introduces the concept of AI co-scientists, which use agentic debate to identify patterns and generate hypotheses, citing an example of identifying a hypothesis about antimicrobial resistance. The post emphasizes the application of AlphaGo's search and reasoning principles to accelerate scientific breakthroughs.
Oct 2025Accelerating discovery with the AI for Math InitiativeThis post announces the AI for Math Initiative, a collaboration between Google DeepMind and five research institutions. It highlights the use of Gemini Deep Think, AlphaEvolve, and AlphaProof as tools to accelerate mathematical discovery. Specifically, it details AlphaEvolve's discovery of a new matrix multiplication algorithm and its application in understanding computational limits. It also mentions Gemini Deep Think's gold-medal performance at the IMO and AlphaProof's silver-medal standard.
Sep 2025Discovering new solutions to century-old problems in fluid dynamicsDeveloped a new AI-driven method for discovering unstable singularities in fluid dynamics equations. This method utilizes Physics-Informed Neural Networks (PINNs) trained to satisfy physical laws, achieving high precision and enabling the systematic identification of new singularity families in the Euler, Incompressible Porous Media (IPM), and Boussinesq equations. The research also identified a pattern in the blow-up speed (lambda) related to the order of instability, suggesting further undiscovered solutions.