The US Department of Energy's Argonne National Laboratory has launched an AI inference service using spare supercomputing capacity to support researchers across DoE labs and the Genesis Mission. The service runs on two clusters: Sophia, with 192 Nvidia A100 GPUs, and Metis, featuring 32 SambaNova SN40L AI accelerators. The platform provides secure access to various large language models, including OpenAI's GPT-OSS, Google's Gemma, and Meta's Llama, through a chatbot-like portal. Researchers are using the service to analyse experimental data in real time, including predicting plasma disruptions in fusion energy research and processing data from particle accelerators and telescopes. Argonne plans to extend the service to its Nvidia GH200-based Tara and B200-based Minerva systems, enabling scientists to experiment with AI without building their own infrastructure.

Researchers at the US Department of Energy's Argonne National Laboratory have developed GridMind, an agentic AI system designed to assist power grid operators through natural language interaction. The system functions as a reasoning co-pilot for control rooms, simplifying complex grid management tasks. GridMind employs a multi-agent architecture where specialised AI agents handle different functions, such as power scheduling and weather-based contingency planning. Large language models coordinate these agents to analyse situations, reason across different tasks and provide explainable recommendations. The system transforms technical analysis into conversational support whilst maintaining rigorous accuracy. Tests on standard power grid models demonstrated that GridMind consistently produced correct results across multiple state-of-the-art language models. The technology aims to accelerate decision-making by integrating disconnected workflows into a coherent reasoning engine.

A research team led by the US Department of Energy's Argonne National Laboratory has developed an AI adviser that optimises machine learning algorithms during autonomous experiments, accelerating discovery of advanced electronic materials. The system was applied to Polybot, Argonne's AI-guided robotic laboratory, to investigate mixed ion-electron conducting polymers for wearable electronics and energy storage. The adviser evaluates algorithm performance in real time and communicates insights to scientists who refine experimental plans. It reduced the study to just 64 experiments from over 4,300 possible combinations. During testing, the adviser suggested switching AI algorithms, leading to significant performance improvements, and identified deposition speed as a key performance driver. The research was published in Nature Chemical Engineering and included collaborators from the University of Chicago, Lawrence Berkeley National Laboratory and other institutions.