Google Research details AI work across scientific fields

Google Research published an overview on May 8, 2025, describing how teams use AI across scientific disciplines. Yossi Matias, vice president and head of Google Research, presents work spanning biomedical science, neuroscience, geospatial science and quantum computing.

By Career Ahead AI — AI-assisted, fact-verified against cited sources. Readers should verify key facts.

The programme spans scientific questions, not one field

Google says AI has long driven scientific advances at the company, and that the pace of breakthroughs today is unprecedented. Its research teams use AI to address fundamental scientific questions across fields including biomedical science, neuroscience, geospatial science and quantum computing. The company also says its teams collaborate with the scientific community across industry and academia. Together, those claims frame the work as a set of research efforts across disciplines, rather than a single application.¹

One example is AI co-scientist, a multi-agent system built on Gemini 2.0. Google describes it as using AI to synthesize information and perform complex reasoning tasks, with the aim of helping scientists create novel hypotheses and research proposals in natural language. The system is presented as support for a particular part of research: forming questions and proposals. That places the emphasis on how scientists may use AI in research work, rather than on a standalone scientific result.¹

A separate project concerns medical diagnostic conversations. Google introduced a multimodal version of AMIE, described as a research AI agent and published in Nature. Google says the system can interpret and reason about visual medical information, working towards more accurate diagnosis. The example brings visual information into the kind of conversational medical research associated with AMIE, while the co-scientist project addresses hypotheses and proposals. They illustrate distinct tasks within the broader scientific programme.¹

That places the emphasis on how scientists may use AI in research work, rather than on a standalone scientific result.1

The neuroscience projects make methods and data central

Google Research and ISTA published the LICONN method in Nature. Google describes it as the first method using commonly available light microscopes to comprehensively map neurons and their connections in brain tissue. The account links the method to access: it says LICONN will enable more labs around the world to pursue connectomics studies. The practical emphasis is on a way of mapping neural connections with commonly available equipment, not only on a new AI system.¹

The other neuroscience example is ZAPBench, the Zebrafish Activity Prediction Benchmark. It includes recordings of more than 70,000 neurons from the entire brain of a larval zebrafish. Google says it open-sourced both the dataset and benchmark to help neuroscientists develop more accurate brain activity models. The project therefore pairs a large neural recording set with a benchmark intended to support the development of models, giving researchers material to work with as well as a way to assess that work.¹

LICONN and ZAPBench serve different research purposes in the examples Google gives. LICONN maps neurons and their connections in brain tissue; ZAPBench supplies recordings of neural activity and a benchmark for model development. Both descriptions put research methods or shared resources alongside AI. For scientists considering these examples, the distinction matters: one project concerns mapping connections, while the other provides recordings and a benchmark for studying activity.¹

Geospatial work connects models with real-world problems

Google describes Geospatial Reasoning as a research effort combining geospatial foundation models with generative AI through a conversational interface. The stated aim is to uncover actionable information. The example adds another kind of work to the list: using a conversational interface to draw on geospatial models and generative AI. It differs from the neuroscience projects, which focus on mapping brain connections or modelling activity, but sits within the same cross-disciplinary programme.¹

Geospatial work connects models with real-world problems Google describes Geospatial Reasoning as a research effort combining geospatial foundation models with generative AI through a conversational interface.

Google also says it launched the first FireSat satellite to help address wildfires. The company describes the satellite as part of a constellation and says the high-resolution imagery is updated globally every 20 minutes. The stated uses include helping emergency responders detect wildfires at an earlier stage and helping scientists understand how wildfires spread. This example connects satellite imagery to both response work and scientific study of wildfire behaviour.¹

Quantum computing appears in two examples. Google describes its Willow chip as a milestone demonstrating error correction and state-of-the-art performance. In a collaboration with Sandia National Laboratories, Google says a quantum algorithm could more efficiently simulate mechanisms needed for sustained fusion reactions. The examples point to different work: one concerns the chip’s performance and error correction; the other concerns simulating mechanisms relevant to fusion. Both are presented as part of the scientific research effort.¹

Choose a research task before judging an AI tool

If you work in scientific research, start by naming the task you need to do before assessing whether an AI system is relevant. Google’s examples range from creating hypotheses and proposals to interpreting visual medical information, mapping neural connections and developing brain activity models. Those are distinct tasks, supported by different systems or resources in the examples. Comparing a tool with a concrete research need follows directly from the range of tasks described.¹

For a research problem involving shared data or methods, check what the project makes available and what it is designed to support. Google says ZAPBench is open-sourced as a dataset and benchmark for neuroscientists developing brain activity models; it describes LICONN as a method for mapping neurons and their connections with commonly available light microscopes. Those details give a researcher a basis for comparing the kind of resource or method at hand with the work they want to carry out.¹

If your work crosses disciplines, examine the collaboration and application described for the project, rather than treating all scientific AI as one category. Google says its teams collaborate with the scientific community across industry and academia, and its examples include wildfire research, medical diagnostic conversations and quantum simulation. A useful next step is to identify which of those research settings resembles your own, then assess the specific task and resource involved. The examples support that comparison without treating the systems as interchangeable.¹

Those details give a researcher a basis for comparing the kind of resource or method at hand with the work they want to carry out.1

Written by Career Ahead AI. AI can make mistakes, so verify key facts. Report errors to support@careeraheadonline.com