Data-driven prediction of piezoresistive response in carbon nanofibers-enhanced cementitious composites using a hybrid deep learning framework

The increasing need for effective Structural Health Monitoring (SHM) systems has accelerated the development of smart cementitious materials with intrinsic sensing capabilities. Among these, self-sensing composites enhanced with carbon nanofibers (CNFs) show significant promise due to their ability to transduce mechanical stimuli into electrical signals. This study presents a hybrid deep learning approach for predicting the fractional change in resistivity (FCR) in self-sensing cementitious composites reinforced, considering the impact of various carbon nanofibers. Building upon previously published experimental results, which systematically evaluated the mechanical, electrical, and piezoresistive properties of CNF-enhanced composites under cyclic compression, two multi-branch neural networks were developed to integrate heterogeneous input modalities. The proposed approach incorporates dynamic loading sequences, static material properties, and formulation-specific categorical inputs processed through dedicated Long Short-Term Memory (LSTM), dense, and embedding layers, respectively. Two model variants were investigated: one relying solely on time-load sequences, and the other incorporating autoregressive FCR information. The dataset was constructed using a sliding window method and partitioned to preserve the distribution of CNF formulations. In the end, a permutation feature importance analysis and ablation study were conducted to quantify the contribution of each input branch and material property to the model performance. Performance evaluation using MAE, RMSE, and R² metrics demonstrated that the model accurately captures the nonlinear resistive behaviour of the composites and generalises effectively across different CNF types and dosages. The results underscore the potential of data-driven frameworks in advancing the predictive capabilities of self-sensing materials for structural health monitoring applications. • Developed hybrid deep learning models to predict FCR in smart composites. • Integrated load sequences, material properties, and CNF types via neural branches. • Compared models with and without autoregressive FCR input for accuracy. • Used sliding window data and feature importance to assess model robustness. • Achieved high prediction accuracy across various CNF types and concentrations.