AI labs and open models supercharge US scientific discovery

By late 2025, the convergence of advanced AI reasoning, robot learning, and lab automation began enabling autonomous laboratories that collaborate with human scientists to accelerate discovery across life sciences, chemistry, and materials science. Oliver Hsu highlighted this emerging synergy, noting that while fully autonomous closed-loop experimentation—where AI independently plans, executes, and iterates experiments without human input—remains a longer-term goal, progress in AI reasoning and physical simulation is uneven but steadily advancing. Early adoption is driven by market demand in sectors like pharmaceuticals and materials science, where mature research output markets justify investment in these novel AI-robotic systems.

Entering 2026, the field shifted from relying on abundant internet data to a data-constrained regime where generating novel experimental data through autonomous labs became essential. Companies like Periodic Labs pioneered integrating AI models directly with physical labs to create closed-loop systems, where reinforcement learning (RL) and large language models (LLMs) iteratively propose, test, and refine hypotheses grounded in real-world experiments. This approach outperforms models trained solely on textual data by using nature itself as the reinforcement learning environment, enabling AI scientists to specialize narrowly but powerfully within domains where reliable reward signals exist.

By mid-2026, the deployment of robotic labs running 24/7 revolutionized experimental throughput, automating repetitive tasks like pipetting and freeing human researchers for higher-level reasoning. OpenAI’s vision exemplifies this closed-loop experimentation cycle: AI models simulate, design, execute, and validate experiments at scale, continuously refining their approaches based on physical feedback—a necessity especially in complex fields like biology where in silico methods alone fall short. This trend is rapidly gaining traction among startups and research institutions, signaling a transformative shift in scientific practice toward autonomous AI-driven discovery.

Large-scale autonomous labs such as Argonne National Laboratory’s RAPID facilities demonstrate the potential to accelerate discovery by tens to hundreds of times compared to traditional methods, conducting thousands of experiments simultaneously—a feat impossible for human scientists alone. Advances in LLMs, chip technology, and robotics are unlocking new scientific capabilities, particularly in materials science and chemistry. Cutting-edge closed-loop AI frameworks integrate vast materials databases, machine learning interatomic potentials, intelligent AI agents, and automated workflows to predict material properties with near atomic-level accuracy, dramatically shortening development times for clean energy technologies despite ongoing challenges in data standardization and experimental validation costs.

The American Science Cloud (AmSC) stands as a foundational AI-integrated high-performance computing infrastructure that unifies the fragmented capabilities of DOE national laboratories into a single, interoperable ecosystem. As Chin Guok, CTO at ESnet, and Berkeley Lab’s Monga emphasize, AmSC leverages ESnet’s high-performance networking and AI-enabled orchestration services to provide scientists with seamless, single-entry access to diverse computational and experimental resources without requiring knowledge of their physical locations. This federated infrastructure prioritizes interoperability over standardization, enabling frictionless reuse of data, compute, and experimental time across grand challenge research areas such as fusion energy and quantum science, supported by DOE’s $600 million Genesis Mission investment.

Argonne National Laboratory exemplifies the transformation of HPC centers into AI-driven scientific collaborators by integrating dedicated AI inference platforms—Sofia, Minerva, and Tara—alongside traditional supercomputers like Aurora. This approach democratizes access to dozens of open-weight AI models through centralized inference services, reducing the need for researchers to manage complex infrastructure themselves. As Uram notes, the combination of AI hardware, inference services, and job submission APIs lays the groundwork for sophisticated AI workflows that not only analyze data in near real-time but also autonomously plan and execute closed-loop simulations, as demonstrated in drug discovery and fusion research.

Oak Ridge National Laboratory’s overhaul of the DOE Atmospheric Radiation Measurement (ARM) program’s computing and storage infrastructure showcases the creation of AI-ready environments that accelerate data access and streamline scientific workflows. The development of the ARM Data Advisor (ADA), an AI agent enabling natural-language interaction for data discovery and personalized recommendations, alongside the Agentic Tooling and LLM Augmentation Stack (ATLAS) framework, exemplifies efforts to standardize AI tool integration and ensure responsible governance. These innovations align closely with the Genesis Mission’s vision of AI-driven scientific discovery by transforming archival data into AI-readable formats and enhancing metadata for trustworthy research.

The DOE’s Genesis Mission is further bolstered by strategic partnerships and investments that expand AI-enabled HPC capabilities beyond national labs. Google’s $40 million commitment, including cloud credits and advanced AI tools like AlphaEvolve and Gemini, empowers tens of thousands of DOE researchers with secure, scalable AI environments that have already demonstrated transformative impacts such as an eightfold reduction in microscope calibration time. Complementing this, Rescale’s integration of agentic AI simulations into DOE supercomputing resources under the Advanced HPC Initiative (AHPI) reduces engineering simulation times by over fivefold, promoting self-service access and ongoing collaboration between national labs and U.S. manufacturers. Meanwhile, Meta’s deployment of open-weight AI models SAM 3 and DINOv3 within DOE’s secure HPC infrastructure slashes scientific image analysis from a month to just 15 minutes, enabling near real-time adaptive workflows that compress discovery timelines dramatically, as highlighted by DOE Under Secretary Darío Gil.

Arcee AI is spearheading U.S. efforts to develop large-scale open-weight AI models tailored for scientific research, exemplified by their 400 billion parameter sparse Mixture of Experts model, Trinity Large, which balances scale and efficiency with 13 billion active parameters. This approach positions Trinity competitively among leading MoE models like MiniMax and GLM 4.5, reflecting a strategic focus on building open scientific AI in public and iterating transparently to foster domestic innovation despite fewer U.S. entities engaging in open model development compared to international counterparts.

The U.S. maintains a commanding lead in AI model development driven by vast compute resources concentrated in companies like Google (25%), Meta (11%), OpenAI (11%), and Anthropic (6%), dwarfing Chinese labs that face resource constraints and conservative safety stances limiting rapid scaling. This advantage has enabled American open-weight models such as Nvidia's Nemotron, Google's Gemma (adopted under the permissive Apache 2.0 license), and Arcee AI’s offerings to stabilize and gain traction within the open model ecosystem, attracting significant developer adoption through permissive licensing and benchmark competitiveness.

Despite the dominance of closed AI models from OpenAI and Anthropic—favored for their economic advantages via query-based monetization—there is growing investor and industry momentum behind U.S.-based open-source or open-weight AI models as strategic alternatives to Chinese AI players. Recent developments include Thinking Machines’ open-weight debut led by former OpenAI CTO Mira Murati, Nvidia’s Nemotron, and OpenAI’s own open-weight releases, signaling an accelerating and volatile AI landscape where open models challenge incumbent dominance and promise broader accessibility and innovation.

Arcee AI’s Genesis-Science-1 (GS1), developed in collaboration with the U.S. Department of Energy and national laboratories, exemplifies a landmark U.S. initiative to create an open-weight AI model specifically designed for reproducible scientific computing. By releasing model weights and tools that empower institutions to operate GS1 within their own infrastructure, the project addresses critical needs for transparency, long-term preservation, and specialized adaptation, while inviting broad contributions from universities, labs, companies, and nonprofits to accelerate its multi-stage development and evaluation.

The DOE Genesis Mission Consortium stands as a pivotal force in the White House's ambitious initiative to double U.S. research productivity through AI, leveraging the American Science Cloud (AmSC) infrastructure and ESnet's high-performance network to unify national labs and facilities into a coordinated, AI-driven scientific ecosystem. With over $600 million allocated to interdisciplinary teams tackling grand challenges like fusion energy and quantum science, the consortium fosters seamless collaboration by integrating Infrastructure Resource Orchestration middleware that enables sophisticated AI workflows across diverse DOE assets.

Expanding its public-private partnership network, the Genesis Mission Consortium has attracted a spectrum of AI startups and technology giants—including Scale AI, Veritone, Wiley, and NASA—each bringing unique capabilities to modernize DOE’s data infrastructure and accelerate discovery. Veritone’s aiWARE platform, for example, transforms fragmented government data into secure, AI-ready assets, while Wiley contributes authoritative, peer-reviewed scientific content and research intelligence tools, underscoring the consortium’s commitment to responsible AI development grounded in expert-validated workflows.

The consortium’s strategic funding and project coordination catalyze breakthroughs across critical domains by integrating vast federal datasets and advanced AI tools from partners like Google and Rescale. Google’s $40 million commitment, featuring AI tools such as AlphaEvolve and Gemini for Government, has already accelerated research productivity—cutting microscope calibration time eightfold at the National Laboratory of the Rockies—while Rescale’s Agentic HPC Pipeline Initiative democratizes access to national lab simulation codes, slashing engineering simulation times by over fivefold and fostering ongoing collaboration beyond traditional DOE-industry partnerships.

Demonstrating a comprehensive and collaborative approach, the Genesis Mission Consortium orchestrates hundreds of AI-powered projects involving over 340 institutions nationwide, backed by substantial funding from the DOE, Department of Defense, and other federal agencies. This includes a $1.5 billion DoD pledge supporting initiatives like the Digital Biosecurity Forge with Lawrence Livermore National Laboratory, which integrates decades of validated defense data to drive AI-enabled breakthroughs in national security and biosecurity, reflecting the consortium’s role in uniting diverse stakeholders to accelerate AI-driven scientific discovery within a secure, sovereign ecosystem.

By mid-2026, the emergence of AI-powered closed-loop research systems marked a transformative leap in accelerating the discovery of advanced energy materials critical to clean energy technologies. These systems, exemplified by the '4th+ paradigm,' seamlessly integrate large materials databases, machine learning interatomic potentials, large language models, intelligent AI agents, and automated laboratory workflows into a unified platform that continuously refines predictions based on experimental feedback. This approach promises to drastically shorten development times for batteries, hydrogen storage, and fuel cells, thereby bolstering the global transition to low-carbon energy, despite ongoing challenges in standardization and model reliability.

Parallel advances in AI-driven scientific imaging have revolutionized materials analysis, with the University of Michigan-led team securing the DOE Genesis Mission Award for pioneering real-time 3D X-ray imaging that compresses analysis timelines from months to near real-time. Leveraging collaborations with Los Alamos and Argonne National Laboratories and harnessing high-performance computing, this initiative accelerates innovation across energy systems, transportation, and national security by enabling rapid insights under extreme conditions.

Meta’s open-weight AI models, SAM 3 and DINOv3, deployed within DOE’s secure national laboratory infrastructure, have slashed scientific image analysis times from roughly a month to just 15 minutes. The SYNAPS-I system, running on 300 NVIDIA A100 GPUs at NERSC, automates complex imaging modalities such as X-ray tomography and neutron scattering, enabling adaptive, automated decision-making that replaces slow manual annotation. DOE Under Secretary Darío Gil highlights that SYNAPS-I 'analyzes data as it’s produced and guides experiments in real time,' exemplifying how AI integration compresses discovery timelines from days to moments and supports the expansive $5 billion Genesis Mission across DOE user facilities.