Abstract
This paper introduces CGMTorch, a novel framework leveraging gradient-based optimization for designing computational granular metamaterials. These materials, composed of granular particles with tunable properties, are optimized for dynamic responses to mechanical vibrations, enabling applications such as acoustic logic gates and waveguides. By utilizing an end-to-end differentiable simulator built on PyTorch, CGMTorch systematically tunes particle properties to achieve desired outcomes in nonlinear and high-dimensional parameter spaces. The framework showcases its potential in user-centric design by simplifying complex configurations and minimizing the computational overhead of offline design optimization for mechanical computation devices. Once the design is optimized, the fabricated passive device operates in real-time without any computational latency, as the computation occurs directly in the material's physical dynamics. This work highlights the transformative potential of granular metamaterials in unconventional computing and user experience enhancement.