Soumyendu Sarkar is a Senior Director and Senior Distinguished Technologist at Hewlett Packard Enterprise and leads Machine Learning research at HPE Labs. His current research focuses on reinforcement learning and agentic LLMs for sustainable data centers, liquid cooling, and energy systems using digital twins. Soumyendu also works on emerging ML areas and is a Principal Investigator on a US DOE-funded AI research project to advance Fusion Energy, as well as HPE’s PI for the Artificial Intelligence Consortium under the National Institute of Standards and Technology. He also leads and participates in several European Union-funded projects on Data Center holistic optimization and Wave Energy. Soumyendu’s other interests include AI for Quantum computing, Tokenomics, Agentic AI, high-assurance ML, trustworthy AI, generative diffusion models, surrogates, and optimization. Before joining HPE, Soumyendu was a Distinguished MTS at Bell Labs and a senior manager on the Intel COE for imaging and computer vision.
Presentation Title:
Toward an Autonomous Data Center Fabric: Physics-Informed Digital Twins, Agentic AI, and High-Assurance Control
Presentation Abstract:
As exascale workloads push data centers to their physical limits, AI must evolve from a passive workload into an autonomous, self-regulating nervous system for the facility itself, addressing stresses in cooling infrastructure, energy grid, and water resources. We need a blueprint for an autonomous data center fabric that replaces static rules with physics-informed digital twins and predictive generative models. At its core, a cooperative Agentic Mesh fuses the reasoning of LLMs with the optimization precision of Multi-Agent Multi-Objective Reinforcement Learning. These agents seamlessly orchestrate deeply coupled infrastructure layers, from liquid-cooled silicon to global scheduling, optimizing performance, energy, and water use. To ensure trust, the fabric is anchored in High-Assurance AI for verifiable logic and bounded safety. This hierarchical control should integrate facilities into the energy grid as active assets, thereby reducing grid stress and coupling power market bidding with real-time physical control of cooling loops, workloads, and battery dispatch.
Coming Soon