Inverse Design of Thermally Active Composite via Policy-Transferred Reinforcement Learning

Data and code of "Inverse Design of Temperature-Responsive 4D Active Composites Using Reinforcement Learning" When you extract the compressed file, you will find the following four files: DATA/├── DATA 1: final_input_data.npy│└── DATA 2: final_output_data.npy DATA: Two NumPy arrays containing the complete dataset used for training and evaluation: final_input_data.npy: (shape: 10000 × 4 × 24): Binary material configurations for the TAC cantilever beam, where each element represents material assignment (0 or 1) in a 4-row × 24-column grid. The dataset comprises 80% randomly generated configurations and 20% patterned designs. final_output_data.npy: (shape: 10000 × 2 × 24): Corresponding deformation trajectories containing x and y coordinates of 24 centerline points along the beam after thermal actuation. CODE/├── CODE 1: s1v1_surrogate.py│├── CODE 2: s2v1_GA_train.py│├── CODE 3: s2v1_SSO_train.py│├── CODE 4: s5v1_1case_DDQN_original_train_nstep.py│├── CODE 5: s6v1_multicase_DDQN_original_earlystop_nstep.py│└── CODE 6: s7v1_multi2single_transfer_DDQN_original_earlystop_train_nstep.py CODE 1: s1v1_surrogate.pyForward prediction surrogate model using neural networks to predict deformation trajectories from material configurations, serving as the environment for RL training. CODE 2: s2v1_GA_train.pyGenetic Algorithm (GA) implementation for inverse design optimization, used as a baseline comparison method. CODE 3: s2v1_SSO_train.pySequential Subdomain Optimization (SSO) implementation that divides the design space into column-wise subdomains for local optimization. CODE 4: s5v1_1case_DDQN_original_train_nstep.pySingle-target Double Deep Q-Network (DDQN) with n-step returns for optimizing material configurations toward individual target shapes. CODE 5: s6v1_multicase_DDQN_original_earlystop_nstep.pyMulti-target DDQN training with early stopping, enabling generalization across diverse target trajectories without retraining. CODE 6: s7v1_multi2single_transfer_DDQN_original_earlystop_train_nstep.pyTransfer learning implementation that fine-tunes the multi-target pretrained model for accelerated single-target optimization, reducing sample usage by up to fivefold.