Mapping Machine Learning to Physics (ML2P)

The Mapping Machine Learning to Physics (ML2P) initiative, spearheaded by the Defense Advanced Research Projects Agency (DARPA), is focused on addressing the critical issue of power consumption in machine learning deployments, especially in resource-constrained battlefield environments. Traditional machine learning systems require significant computational power, a challenge magnified when these systems are deployed at the edge. ML2P aims to tackle this problem by making power efficiency a fundamental design consideration, using precise Joule-based measurements to directly correlate ML system efficiency with underlying physical constraints. This approach enables more reliable predictions of power usage and performance across a broad range of hardware platforms. The program intends to develop advanced multi-objective optimization methods that seek an optimal balance between power consumption and performance. In addition, it will introduce and utilize the concept of Energy Semantics of ML (ES-ML) to understand how localized optimizations affect overall energy efficiency in machine learning applications. The contract solicitation is open under announcement number DARPA-PS-25-32, with responses due by December 17, 2025. The project underscores the Department of Defense’s commitment to pushing the frontiers of energy-aware machine learning, dedicated to solutions that are not only high-performing but also power-efficient, which is crucial for reliable ML operation in austere environments.