MURI AI-Guided Self-Organization: Tailoring Disorder to Shape Complex Nonlinear Dynamics (MURI Topic #4)Tracking Number 23-000005209
Abstract
Project Abstract (Approved for Public Release)Research problem: Work spanning three decades has shown that perturbative, structuraldeformations can control self-organization in complex nonlinear systems. There has, however, been little progress generalizing such"control through disorder" beyond idealized models. Machine learning is a promising tool to make this step, tackling the complexityand opaqueness of real experiments. AI-based control has achieved superhuman performance in video games, but hasyet to be deployed to high-dimensional physical experiments. Objectives and approaches: We will combine modern machine learning with physical insightsand high-throughput automated experiments to enable unprecedented AI-assisted design and control of complex self-organization throughdisorder. Guided by studies with universal physical models, our team will develop new techniques for physics-guided inverse design and control. We will develop experiments spanning multimode fiber lasers, semiconductor laser arrays, and networks of electronic oscillators and RF SQUIDs. These will function jointly as a training andtesting ground to enable universal techniques for AI-based design and control of complex systems (i.e., an "ImageNet for AI-control").Research outcome: Enabled by unique approaches to gradient-based learning and highthroughput experiments, our team will deploy AI-based design and control to tasks with action spaces of 1000 times higher dimensionality than action spaces so far considered in reinforcement learning. High dimensionality is a frontier for bothAI control and nonlinear physics, and will require disorder-like high-dimensional "smart symmetry breaking." We will also use complex nonlinear systems themselves as effectively unlimited bandwidth, closed-loop neural network controllers, allowing neural network control to be applied to ultrafast systems such as semiconductor laser arrays.Research team: Our team s strength is multifaceted, complementary diversity of expertise, ideas, and experimental platforms. We combine a long history of research underpinning MURI Topic4 with modern techniques and unconventional ideas that fuse physics with AI. Hui Cao (Yale) has pioneered disorder-enhanced laser technologies, including high-power semiconductor lasers and random lasers. Herbert Winful (Michigan) made foundational contributions to the physics of coupled lasers. Ying-Cheng Lai (Arizona State) and Vassilios Kovanis (Virginia Tech) made seminal contributions tocontrol and machine-learning of dynamical systems. Tsampikos Kottos (Wesleyan) and Kovanis pioneered the use of disorder as a resource to synchronize oscillators and lasers, and Kottos is an expert on non-Hermitian physics. Steven Anlage (Maryland) is a leader insuperconducting microwave devices and control of complex scattering with machine learning. Logan Wright (Yale) conducted the largest-ever laser synchronization experiments and pioneered techniques for efficient high-dimensionaldeep-learning design of physical systems. DoD Impact: Our focus will be on basic research, discovering new physical concepts and AI techniques that can universally control complex nonlinear dynamics and self-organization. As a culmination of our team s efforts, we will realize a universal algorithm for controlling and designing physical systems, and release diverse datasets to facilitate further development. This algorithm wouldbe transformative for many DoD engineering and logistical challenges involving complex physical systems, including high-power energy delivery systems, robotics, manufacturing, autonomous space and submarine systems, and control of high-energy fluids and plasmas. We will devote special attention to educating the next generation of military scientists, and, as demonstrations of our AI-enabled control and design, to realizing reconfigurable phaselocking in large 2D semiconductor laser arrays and power-scalable mode-locking in fiber lasers.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Nov 08, 2024
- Source ID
- N000142412548
Entities
People
- Hui Cao
Organizations
- Office of Naval Research
- United States Navy
- Yale University