Active Learning for Data‐Efficient Optimization of Entangled Granular Metamaterials

Abstract

Entangled granular materials derive exceptional macroscopic rigidity from topological interlocking of their constituents. However, rationally designing particle geometries to maximize this effect remains a formidable challenge due to the immense configuration space and the prohibitive computational cost of discrete element simulations. Here, we present a data‐efficient and uncertainty‐aware design framework for autonomous discovery of optimal particle geometries that substantially enhance macroscopic yield strength beyond conventional heuristic shape families. A Semi‐Supervised Graph Wasserstein Autoencoder embeds discrete particle topologies into a continuous latent manifold, while Langevin‐based stochastic optimization explores the structure‐property landscape. By explicitly penalizing epistemic uncertainty through a Lower Confidence Bound acquisition strategy, the framework robustly avoids spurious optima under extreme data scarcity. The discovered low‐aspect‐ratio particle geometries achieve quasi‐cohesive behavior, high yield strength, and self‐supporting capability typically associated with slender fibers. Compression and angle‐of‐repose experiments on 3D‐printed assemblies confirm high simulation fidelity and validate geometry enhanced, interlocking‐dominated load‐bearing architectures.