This study presents a data-driven framework integrating COMSOL Multiphysics automated simulations with machine learning, generative modeling, and reinforcement learning (RL) to design tetrahedral pentamode (PM) unit cells that mimic water‘s acoustic properties while ensuring manufacturability. The framework continuously generates diverse geometric configurations, conducts automated simulations, and analyzes essential physical properties, including bulk modulus, shear modulus, density, and impedance. A Conditional Variational Autoencoder (CVAE), embedded within the RL framework employing Proximal Policy Optimization (PPO), predicts relationships between geometric parameters, e.g., cone radius, height, angle, and corresponding physical properties. Custom reward functions guide PPO agents through dynamic training episodes, targeting specific design objectives such as maximizing the bulk-to-shear modulus (B/G) ratio. By dynamically adapting learning rates and leveraging predictive capabilities of the CVAE’s latent space, the framework effectively explored the PM design space. Generated unit cells were validated using COMSOL simulations, confirming that they consistently met targeted design criteria and achieved a notable B/G ratio improvement of about 170. We explored Conditional Wasserstein GANs with Gradient Penalty (cWGAN-GP), augmented by CoordConv layers to enhance generative capabilities. The cWGAN-GP demonstrated advantages in spatial feature representation and direct physical to image synthesis.
Qiu et al. (Wed,) studied this question.