Mapping Machine Learning to Physics (ML2P)

The Defense Advanced Research Projects Agency (DARPA), through its Information Innovation Office (I2O), has released the Mapping Machine Learning to Physics (ML2P) program solicitation under funding opportunity number DARPA-PS-25-32. This initiative reflects DARPA’s long-standing mission to advance technological frontiers critical to national security. The ML2P program seeks to address a major challenge in machine learning (ML): its growing demand for energy. As ML systems scale in size and complexity, their energy requirements increase sharply. For the Department of Defense (DoD), which often operates in energy-constrained environments such as on the battlefield, more efficient use of power is an operational imperative. The objective of ML2P is to fundamentally reframe power consumption as a first-class consideration throughout the machine learning lifecycle. The program will create methods to preserve local energy semantics and establish tunable energy-performance functions. By enabling accurate prediction of ML models’ power and performance trade-offs, ML2P will lay the groundwork for more efficient hardware design and software optimization. This includes exploring new algorithms, developing multi-objective optimization functions, and capturing energy semantics to inform future model construction. A key transition goal is to make ML2P outputs open source, accessible through major repositories such as scikit-learn and DARPA’s GitHub page, ensuring broad academic and commercial utility. The program is structured into two phases of 12 months each, with a go/no-go decision point at the end of Phase 1. The first six months focus on experimental setup, hardware procurement, objective function selection, and development of energy semantics for ML (ES-ML). Subsequent work will refine algorithms, optimize multi-objective functions, and produce software capable of generating energy-aware ML models optimized across accuracy and efficiency. Test and evaluation will be conducted independently by a government-selected team, ensuring robustness and reproducibility of results. DARPA anticipates awarding multiple Other Transaction (OT) agreements for prototype projects, leveraging flexibility outside the traditional Federal Acquisition Regulations framework. A total of approximately $5.9 million is available, divided across multiple performers, with $3.5 million planned for Phase 1 and $2.4 million for Phase 2. Cost sharing may be required under 10 U.S.C. § 4022, especially for traditional defense contractors not partnered with nontraditional entities. Deliverables include monthly reports, annotated presentations, open-source software and documentation, technical papers, and system development plans. Eligibility for ML2P is open to a broad range of applicants, including large and small businesses, nontraditional defense contractors, research institutions, and universities. However, UARCs, FFRDCs, and entities with conflicts of interest are discouraged from applying unless they receive prior approval. Proposals must represent unclassified, fundamental research and avoid controlled unclassified information (CUI). Proposers are expected to demonstrate technical innovation, feasibility, transition potential, and cost reasonableness. The selection process involves an abstract submission due October 6, 2025, followed by government invitation to submit full proposals and participate in oral presentations beginning December 15, 2025. The timeline includes key milestones: proposers’ day on August 26, 2025; question deadline on October 3, 2025; abstract deadline on October 6, 2025; and proposal submission by December 8, 2025. Selected projects will launch with a kickoff meeting in early 2026, followed by principal investigator meetings scheduled through fiscal years 2026–2028. Oral presentations will be conducted by invitation only. The program emphasizes collaboration, reproducibility, and open dissemination of research findings, reflecting DARPA’s strategy to accelerate energy-efficient ML innovation for defense and beyond.

Proposers should specify open-source licenses without restrictions; restrictive licensing will be a significant weakness; budgets must match technical effort rather than targeting ceiling amounts