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

Area of Science:

• Materials Science and Engineering
• Artificial Intelligence and Robotics

Background:

• Active composites (ACs) offer autonomous shape transformation for applications in soft robotics and biomedical devices.
• Inverse design of multi-material 3D printed ACs is computationally challenging due to vast combinatorial possibilities.

Purpose of the Study:

• To develop a reinforcement learning (RL)-based framework for the inverse design of thermally active composites (TACs).
• To enable rapid and accurate design of TACs for complex shape transformations, minimizing deformation errors.

Main Methods:

• Reformulated inverse design as a sequential decision-making process using a 4x24 grid decomposition.
• Employed a target-conditioned RL policy trained on multiple target trajectories for versatile shape generation.
• Validated performance against genetic algorithms (GA) and sequential subdomain optimization (SSO) based on sample count and function evaluations (FEs).

Main Results:

• The RL framework achieved high target accuracy (RMSE ≤ 0.1) with significantly fewer samples and FEs compared to GA and SSO.
• Experimental validation using 4D-printed TAC specimens demonstrated successful replication of complex, free-form trajectories.
• The learned policy showed effective transferability for accelerating single-target optimization.

Conclusions:

• Reinforcement learning agents can efficiently perform sequential material design through long-term reward optimization.
• The proposed framework offers a powerful tool for intelligent design and manufacturing pipelines of active materials.
• Demonstrates potential for creating complex, functional structures with precise shape-shifting capabilities.