Inverse design of thermally active composite via policy-transferred reinforcement learning

Active composites (ACs) capable of autonomous shape transformation under external stimuli enable new opportunities for soft robotics, biomedical devices, and intelligent structures. However, the combinatorial design space of multi-material 3D printing makes inverse design computationally intractable. Here, a reinforcement learning (RL)-based framework is proposed that reformulates the inverse design of thermally active composites (TACs) as a sequential decision-making process. A 4 × 24 grid is decomposed into 24 column-wise decisions to minimize deformation error with respect to target trajectories. A single target design was first demonstrated for an individual trajectory. A target-conditioned policy was then learned using multiple targets to enable rapid design across diverse shapes. The multiple target policy was further transferred to accelerate single target optimization. Performance was evaluated against genetic algorithm (GA) and sequential subdomain optimization (SSO) using the number of samples and function evaluations (FEs) under identical termination criteria. Experimental validation was conducted using 4D-printed TAC specimens via grayscale digital light processing (g-DLP), and demonstrations with complex trajectories, including free-form KAIST logo patterns, confirm that the proposed framework achieves target accuracy (root mean square error ≤ 0.1) with low samples and FEs. This study demonstrates that an RL agent can rapidly perform sequential material design through long-term reward optimization, indicating its potential for extension to intelligent design and manufacturing pipelines.