In-memory phononic learning toward cognitive mechanical intelligence

Mechanical intelligence seeks autonomous materials that embody memory, perception, learning, and decision-making within the mechanical domain. A fundamental challenge is to integrate memory with information processing in an efficient, scalable, and interpretable way that supports cognitive learning and decision-making. Here, we introduce in-memory phononic learning, a framework that unifies non-volatile mechanical memory with wave-based perception in a phononic metastructure. Spatial information is stored as structural states that program elastic-wave responses, allowing probing waves to act as queries that retrieve informative features through parallel, physically transparent localization modes. Learning is achieved by offline optimization of input waveforms to selectively combine these features for multitask memory-pattern recognition without hardware redesign, with decisions inferred directly from output-energy responses. We experimentally and numerically demonstrate high-accuracy classification across tasks from geometric shapes to handwritten digits. Overall, this work elevates “materials that compute” toward “cognitive matter” that can interpret, reason, and learn through wave physics, laying a foundation for scalable and robust autonomous material systems.