Google DeepMind、米エネルギー省の科学加速計画Genesisを支援 Google DeepMind supports U.S. Department of Energy on Genesis: a national mission to accelerate innovation and scientific discovery

Google DeepMind has announced its participation in Genesis, a national initiative led by the U.S. Department of Energy (DOE) to accelerate scientific discovery by tightly coupling frontier AI with the DOE's world-class supercomputing infrastructure and decades of scientific data.

The DOE's seventeen national laboratories operate some of the most capable HPC systems on the planet, including the Frontier exascale machine at Oak Ridge and Aurora at Argonne, and they curate vast datasets across fusion energy, high-energy physics, chemistry, materials, climate and biology. Genesis aims to weave these computational and experimental assets together with modern AI — both general-purpose foundation models and science-specific systems — to compress the loop from hypothesis generation, to simulation, to experimental design and validation.

Google DeepMind brings a notable track record in AI for science. AlphaFold transformed structural biology, GNoME proposed millions of candidate inorganic materials, WeatherNext pushed medium-range forecasting, and the lab has published work on controlling tokamak plasmas with reinforcement learning. Combined with the broader Gemini model family, these capabilities could plausibly help national-lab researchers triage candidates, steer expensive simulations, and surface patterns hidden in petabyte-scale datasets, though the precise scope of DeepMind's contribution to Genesis has yet to be detailed.

The announcement fits a wider U.S. policy push to treat AI for science as a strategic priority. It echoes the White House's AI Action Plan and the DOE's emerging FASST framework — Frontiers in AI for Science, Security and Technology — which envisions a public-private compute and model stack purpose-built for research. DOE labs have similarly engaged Anthropic, OpenAI, NVIDIA and others, signaling that frontier-AI partnerships are becoming a standard component of national science infrastructure rather than one-off pilots.

There are real questions to watch. Sensitive datasets, export-control considerations, and the need for reproducibility and verifiability in scientific AI all complicate deployment; foundation models that hallucinate are a poor fit for nuclear, biosecurity or materials work without careful guardrails. It is also unclear how broadly any Genesis-derived models or tools will be released to the wider research community versus kept inside lab perimeters.

Still, the direction is significant. If Genesis manages to make frontier models a routine instrument alongside beamlines, telescopes and supercomputers, it could meaningfully shorten the cycle time of discovery in areas the U.S. has flagged as strategic — fusion, advanced materials, batteries, semiconductors and biotechnology. Whether that translates into measurable gains will depend less on model scale than on data plumbing, evaluation rigor, and how effectively scientists and AI teams co-design workflows over the coming years.