Artificial Intelligence

The U.S. Department of Energy’s (DOE’s) Atmospheric Radiation Measurement (ARM) User Facility is systematically integrating AI into its workflows to enable more efficient data discovery, robust quality control, and the production of advanced, value‑added data sets. ARM’s AI plan supports DOE’s sweeping national AI initiative, the Genesis Mission.

When applied to atmospheric research challenges, AI has the potential to accelerate discovery from ARM’s observations and data products.

A strong foundation of standards, governance, and ethical guidelines is critical for scaling AI across ARM in a sustainable and responsible way. ARM created the AI in ARM Team to develop a governance document for AI and support AI activities in ARM. The AI governance plan will ensure that ARM’s AI efforts are transparent, reproducible, and aligned with DOE’s goals for trustworthy AI practices.

As part of this effort, ARM will establish consistent standards for evaluating its AI models before they are used in operations. ARM will work to define AI and machine learning (ML) standards within Data Object Design documentation and produce an ML best‑practice guide for ARM developers.

ARM users and staff are encouraged to review AI-generated content for factual accuracy and clarity to remain in alignment with scientific integrity standards. In addition, ARM will create example AI disclosure templates to help users and staff preserve the integrity of images and scientific figures. Together, these efforts will provide the guidance and technical foundation needed to responsibly integrate AI into ARM’s platforms while maintaining alignment with DOE-wide priorities.

ARM is currently developing an AI-ready infrastructure that can scale with the needs of the ARM user community while maintaining compatibility with the American Science Cloud, a secure, connected and integrated, science-optimized environment that links DOE’s computing and experimental facilities. The approach will incorporate continuous learning pipelines that enable incremental updates and create progressively refined AI models that adapt to evolving project requirements across atmospheric science applications. Key components of this infrastructure include:

• Computational infrastructure upgrades: Enhancements are being made to support the demands of AI and scientific applications. This includes adding graphics processing units (GPUs) to ARM’s Cumulus supercomputer to boost performance for ML and earth system modeling. New GPUs are also being installed that support large language models (LLMs), which can generate text, code, summaries, or other outputs in response to user prompts. This hardware enhancement will support “embedding generation,” a technique for transforming data into numeric representations that help AI systems identify meaning and relationships. It will also support vision-based workflows to help analyze images, videos, and documents alongside text. A new AI-optimized file server has been introduced, connecting directly to the GPU environment and providing faster access to ARM data and metadata. In parallel, ARM is expanding storage capacity and exploring more efficient ways to store and search complex data sets used in AI systems, such as document representations, metadata, and scientific information, enabling users to access the information they need much more quickly.
• Coming soon—The ARM Data Advisor (ADA, pronounced ā-duh), a conversational AI assistant specializing in ARM data, offers a simple, intuitive interface designed to guide users through the process of data discovery and access. ADA can identify data sets relevant to the user, offer context as to why those data sets are relevant, and allow the user to place a data order without leaving the advisor interface. Through ADA’s integrated feedback feature, users can rate its responses, enabling continuous improvements over time.
• Coming soon—The ARM Agentic Tooling and LLM Augmentation Stack (ATLAS) serves as a cohesive application framework that allows AI-enabled scientific workflows across multiple ARM platforms, allowing a scientific user to ask a question once. ATLAS is the foundational framework that will be available alongside ADA, which is the primary application built on top of ATLAS. Multiple agents collaborate to deliver a curated, summarized view of information that may include metadata, instrument information, data products, documentation, publications, and other supporting scientific knowledge. ATLAS leverages ARM Agentic Retrieval-Augmented Generation (A-RAG), an ARM-specific flavor of RAG that goes beyond a simple, single retrieval task; rather, multiple specific agents work together to perform distributed retrieval, synthesis, and composition of pertinent information from the ARM infrastructure, the web, and scientific documentation. Unlike existing RAG systems, which use a single retrieval path, A-RAG enables coordinated reasoning among multiple agents and access to both tools and domain-specific information to support contextually rich scientific workflows. The goal of ATLAS is for multiple agents to collaboratively retrieve, synthesize, and present the most relevant information needed to support a user’s scientific workflow or research question.
• Coming soon—The Ask ARM search assistant will provide ARM.gov users with reliable, source-grounded answers to questions, pulling information from multiple ARM platforms. Ask ARM will respond quickly to user queries, enhancing website navigation, reducing search time, and improving the overall ARM.gov user experience.
• A model storage solution is planned to ensure trained models maintain full reproducibility with tracked hyperparameters (the configuration variables set before training a model). Once implemented, the model storage solution will be made available for systematic reuse across research teams.

ARM is actively applying AI to ARM’s extensive observational data sets to develop enhanced data products for scientific discovery. Existing AI-enabled data products use ML to do the following:

• produce cloud masks from ARM micropulse lidar data
• provide additional quality checks on merged aerosol size distributions from ARM’s scanning mobility particle sizer and aerodynamic particle sizer
• improve current algorithms for classifying synoptic weather regimes and best estimate planetary boundary-layer height, which rely on threshold-based methods
• estimate cloud cover and thickness, detect artifacts (e.g., bugs and dirt), and report uncertainties in ARM all sky imager data.

ARM’s AI-enabled data products are identified in metadata found on ARM’s Data Discovery browser, on data product web pages, and in technical reports. AI-enabled versions of previously existing products have “machine learning” in their name to differentiate between the AI and non-AI versions.

Current AI-enabled data product development activities within ARM include the automation and integration of data epochs—periods of well-characterized, calibrated measurements focused on particular atmospheric phenomena—within Data Discovery. ARM is also working on a value-added product for classifying thermodynamic cloud phase using ML models.

ARM’s next focus is to transition these ongoing development efforts into sustained, production-level data products that can be broadly used by the community, reducing barriers to adoption and accelerating discovery. ARM will also continue to explore the development of new AI-enabled data products and tools, such as applying AI and edge computing to process Doppler lidar spectra and extract additional information from the data.

ARM is committed to advancing the use of AI within the atmospheric research community and improving the accessibility of ARM data. A working group of ARM staff supports collaboration and implementation while an external advisory group provides guidance to ARM on AI-related issues. Planned engagement activities such as webinars, workshops, summer schools, hackathons, and AI‑focused sessions at major ARM meetings ensure that community needs and feedback directly inform AI priorities. These efforts are designed to serve both researchers already applying AI methods, who will benefit from greater visibility into ARM’s AI activities, and the broader ARM user community. If you have questions, are interested in activities, or need AI resources, contact the AI in ARM Team.