How artificial intelligence is improving scientific research for practical benefit.

Geospatial reasoning gives intelligent maps a voice and cognitive process. You can ask it simple questions like a friend.

Google research teams employ AI to answer basic scientific questions and study quantum computing, geospatial science, biological science, and neurology.

Artificial intelligence has historically driven scientific progress at Google, and the current rate is unmatched. AI's advancements have accelerated and expanded the "magic cycle" of research from breakthrough to practical impact.

AI boosts human creativity. Google teams are using AI to answer basic scientific questions and broaden the range of possibility, yielding new insights into life and innovative solutions to humanity's biggest issues. To accelerate scientific discoveries, collaborate with academics and industry. It also gives partners all technologies and instruments for study.

Recent Google Research findings in these four areas have major scientific and social ramifications.

Improving disease treatment with biological science

AI's promise to democratise science, personalise medicine, and expand biological and medical research fascinates us. This AI co-scientist wants to accelerate biomedical therapy discovery. This multi-agent system uses AI's ability to synthesise information and perform complex reasoning to help scientists produce fresh research ideas and hypotheses using plain language.

A multimodal version of AMIE, a medical diagnostic dialogue AI agent, was published in Nature. It can interpret visual medical data for more accurate diagnosis. AMIE is based on MedPaLM and other medical language models.

Using embedding models for digital pathology, dermatology, and chest x-rays, the developers created TxGemma, a set of open models to improve therapeutic development. To help developers build medical AI apps, it keeps supporting Health AI Developer Foundations.

Genomic research is also used to diagnose rare disorders and investigate genetic susceptibilities. Researchers can correlate genomic variants with REGLE, an unsupervised deep learning algorithm. In collaboration with Personalised Pangenome References, researchers released new DeepVariant models. When analysing genomes with many ancestries, these models reduce errors by 30%.

Neuroscience research is improving brain understanding

Have also progressed brain science and connectomics in the previous decade. The first method for mapping neurones and their connections in brain tissue using commonly available light microscopes, LICONN, was published in Nature yesterday by Google Research and ISTA. LICONN will enable connect omics research in more labs worldwide.

In collaboration with Harvard and HHMI Janelia, the developers released the Zebrafish Activity Prediction Benchmark (ZAPBench) beyond neural connections. Over 70,000 larval zebrafish brain neurones were recorded for this benchmark. Scientists can now study the relationship between dynamic neural activity and structural circuitry in a vertebrate brain. The dataset and benchmark are publicly available to help neuroscientists model brain activity more accurately.

In collaboration with Princeton University, NYU, and HUJI, investigations examined the similarities and differences between deep language models and the human brain in natural language processing. This study suggests that deep learning models may offer a new computational framework for brain neuronal code decipherment.

Addressing global concerns via geospatial reasoning

Google Research speeds up geographical problem-solving by making important information more accessible. The first FireSat satellite was launched to fight wildfires. The high-resolution data, updated globally every 20 minutes to build the constellation to over 50 satellites, will help scientists comprehend wildfire propagation and emergency responders notice flames early. Companies have improved climate resilience and crisis response with Flood Forecasting and WeatherNext AI models.

The new Geospatial Reasoning research project uses generative AI and geospatial foundation models to locate actionable knowledge using a conversational interface. It builds on prior Population Dynamics and trajectory-based mobility core models, as well as weather, floods, wildfires, Open Buildings, and SKAI models. Geospatial reasoning benefits public health, integrated business planning, urban planning, climate research, and others.

Quantum computing nears practicality

For over ten years, it has been working towards constructing huge quantum computers that can solve impossible problems. This Willow chip, with mistake correction and cutting-edge performance, is a milestone. On World Quantum Day, stressed its progress towards practical applications.

In collaboration with Sandia National Laboratories, researchers found that a quantum algorithm might better simulate sustained fusion processes. This could enable fusion energy with its large-scale clean energy potential. It demonstrated a breakthrough hybrid method to quantum simulation that opens the door to future scientific discoveries that will advance quantum research.