Condensed Matter Seminar: Learning Dynamical Behaviors in Mechanical Systems

Materials and other complex systems are typically composed of vast numbers of microscopic degrees of freedom, and modifying these opens the door to novel functionality. Traditionally, achieving a desired function requires designing and then modifying the system’s microstructure accordingly. In contrast, recent ideas inspired by learning theory have introduced local evolution laws for the microstructure that, if followed, allow a material to autonomously evolve toward a target function, effectively enabling the material to solve its own design problem. Such local learning rules have been proposed as a means to tune material properties, as a physical platform for efficient AI computation, and even as a potential mechanism for learning in living systems.

While prior work has mostly focused on the quasistatic regime, this talk explores the dynamical regime, governed by damped Newtonian dynamics. We will discuss key considerations for controlling dynamics and introduce several learning rules. Finally, we will illustrate how these principles allow simple mechanical or electronic systems to develop complex functionalities, such as classifying temporal signals, recognizing spoken words, and adjusting the rate dependent response of materials.