Energy-aware machine learning

Traditional machine learning models are designed with a single focus: maximizing performance.

While this approach has driven breakthroughs in language models, image recognition, and other domains, it has overlooked a critical factor — electricity consumption.

Our Mapping Machine Learning to Physics (ML2P) program is designed to transform how artificial intelligence (AI) systems balance performance with energy use.

ML2P addresses this gap by mapping machine learning model performance to physical electric characteristics, with a focus on measuring energy use in joules. By embedding energy-awareness into the design of AI systems, ML2P aims to create models that can achieve the right balance between accuracy and power consumption.

“In an era where AI is increasingly deployed in power-constrained environments, such as at the tactical edge, energy efficiency is no longer optional,” said Bernard McShea, founding program manager for ML2P. “With ML2P, we want to move beyond optimizing just for accuracy and instead understand, for every joule of electricity, what level of performance we’re getting back. That will enable us to build AI that is smarter, leaner, and more useful to the warfighter.”

The program is calling on experts from across disciplines — including electrical engineering, mathematics, logic, and machine learning — to help design a new generation of “energy-aware” machine learning (ML). ML2P will guide model construction by continuously considering all prior design choices and creating training functions that optimize the trade-off between energy consumption and model performance. Together, these tools enable the principled and computationally efficient development of energy-aware machine learning models.

McShea says ML2P’s impact could extend well beyond defense applications. The program could make large language models, generative AI, and classification tasks more energy-efficient while maintaining their performance on existing hardware. Beyond immediate efficiency gains, ML2P could also contribute to the design of better hardware in the future.

“Making models more efficient and performant is crucial as AI applications often require substantial computational resources, leading to high energy consumption,” he said. “By enabling principled simulation of machine learning model performance on general-purpose compute systems, it could provide insights into how hardware should be optimized for AI workloads.”

If successful, ML2P will establish a new paradigm for AI design: one where efficiency and performance go hand in hand, ensuring that AI systems can economically operate at speed and scale with high levels of accuracy.

Researchers interested in learning more about ML2P and/or submitting a proposal may visit the program solicitation at SAM.gov. A link to the replay of the ML2P Proposers Day can be found on the ML2P program page.

###

Media with inquiries should contact DARPA Public Affairs.