CPS: Medium: Collaborative Research: A Function Encoder Framework for Transferable and Adaptive Cyberphysical Intelligence in Metal Additive Manufacturing

Modern cyber-physical systems, including advanced manufacturing platforms, robots, and autonomous systems, often operate in dynamic and rapidly changing environments. Yet existing approaches often require extensive redesign, retraining, or manual tuning when operating conditions change. This project develops learning-enabled control systems that can generalize knowledge across tasks, adapt to changing operating conditions in real time, and integrate physical principles that govern the systems they control. These capabilities are essential for creating technologies that are reliable, adapt to changing conditions, and make effective decisions with limited data. While motivated by emerging robotic and hybrid manufacturing systems that combine additive and subtractive operations, the resulting methods are broadly applicable across diverse cyber-physical domains. The project will advance fundamental understanding of how intelligent systems can learn and generalize in complex environments, while supporting workforce development through research training, curriculum development, and STEM outreach. These outcomes will strengthen U.S. leadership in advanced manufacturing and intelligent systems, contributing to economic competitiveness, technological innovation, and national security. This project develops a unified framework for transfer learning, real-time adaptation, and physics-informed modeling in cyberphysical systems. The central innovation is the use of function encoders, neural representations that learn compact function-space bases reusable across tasks, environments, and system configurations. The research effort consists of three thrusts. The first develops foundations for transfer learning using function-space representations, including generalization guarantees and uncertainty quantification. The second develops basis-to-basis operator learning methods that formulate adaptation as operator inversion in Hilbert spaces, enabling rapid inference of latent system parameters and uncertainty-aware feedback control. The third incorporates physical laws and multi-fidelity models through physics-consistent basis functions and structured mappings between models of different fidelity. The methods will be deployed and validated on next-generation robotic manufacturing systems, including multi-robot hybrid manufacturing platforms that integrate additive and subtractive operations and robotic wire-arc additive manufacturing systems. The anticipated outcome is a new class of transferable, adaptive, and physics-informed learning methods that advance the foundations of cyberphysical systems and learning-enabled control. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.