Directional Latent Hybridization: Beyond Random Noise in Physics‐Informed Generative Inverse Design of Nonlinear Metamaterials
Abstract
Conventional metamaterial inverse design is hindered by stochastic generative models that lack physical grounding and often fail to reach optimal performance. Here, we introduce a framework coupling a physics‐informed neural operator (DeepONet) and a 3D geometry generator (3D‐cVAE) with a Directional Latent Hybridization (DLH) strategy. By merging dominant traits from parent geometries in the latent space, our approach enables deterministic latent optimization, yielding a high prediction accuracy for energy ( R 2 = 0.957) and force ( R 2 = 0.981). Unlike random noise‐based generation, which suffers from a 0% success rate at high volume fractions ( V r = 0.8), the DLH strategy maintains a 73% success rate and achieves significantly lower volume errors (< 4.81%). Experimental validation using additively manufactured thermoplastic polyurethane (TPU) lattices confirms that DLH‐optimized architectures exceed the performance of their base designs, achieving up to 24.03 J of absorbed energy and a peak force of 9.80 kN. This framework establishes a novel physics‐informed generative paradigm for discovering next‐generation metamaterials with physically interpretable and predictable performance.