Multi-Fidelity Transfer Learning for Defense Applications

SBIR topic MDA27BZ02-DV010 (U.S. Department of Defense – Missile Defense Agency), Phase I and II. Open to U.S. small businesses. Opens: December 2, 2026. Closes: January 6, 2027. Rapid Artificial Intelligence (AI) integration into the MDS is required for the United States to properly defend itself against its adversaries. The development of robust Artificial Intelligence/Machine Learning (AI/ML) agents for complex missile defense and aerospace applications is often constrained by the computational cost, limited availability, and security restrictions of high-fidelity simulation environments. Transfer learning provides a critical pathway to overcoming these barriers. By initially training agents on low-fidelity models or unclassified representations—where data is abundant and computational demands are reduced—core behaviors, strategies, and decision-making frameworks could be efficiently developed and refined. These pretrained models could then be adapted and optimized in high-fidelity and/or environments at higher classification levels, allowing the agent to leverage its foundational knowledge while rapidly acquiring the nuanced performance required for operational deployment. This staged approach significantly reduces training time, minimizes resource expenditure, and mitigates the risk of overfitting to scarce high-fidelity/classified data. Moreover, it enables iterative development cycles where low-fidelity environments serve as agile testbeds for innovation, and high-fidelity environments refine and validate for mission-readiness. Leveraging transfer learning in this manner not only accelerates capability delivery but also ensures that AI/ML systems achieve the precision, resilience, and adaptability necessary for real-world mission success. Official solicitation: https://www.dodsbirsttr.mil/topics-app/