Reverse Algorithm for Mechanical Properties of Thin-Film Material Based on CNN-LSTM Neural Network

This study proposes a hybrid CNN-LSTM deep neural network for reversing constitutive parameters from nanoindentation load-penetration depth (P-h) curves. A dataset comprising 125 numerically P-h curves with combinatorially sampled Johnson-Cook parameters (A, B, n) was generated via finite element (FE) simulations. Network architecture integrates CNN (2 layers) to extract local morphological features and LSTM (4 layers) to capture temporal dependencies, enabling spatiotemporal joint learning of indentation responses. Hyperparameter optimization was conducted in two stages: an initial determination of CNN/LSTM layers based on test-set mean absolute percentage error (MAPE), followed by Bayesian optimization iterations to finalize the optimal configuration. Comparative experiments demonstrate that the CNN-LSTM model significantly outperforms standalone counterparts, achieving a 5 4% reduction in overall MAPE (1 4. 9 0%) compared to the CNN model (32.33 %) and a 24 % reduction compared to the LSTM model (19.47 %). Notably, consistent error reduction trends are observed across all individual parameter predictions. This approach substantially reduces reliance on iterative FE simulations, exhibiting strong generalizability and industrial applicability for rapid mechanical characterization.