LSTM architecture ablation for cross mix generalization on full flexural curve forecasting of sustainable hybrid macro basalt fiber reinforced UHPC

Learning nonlinear mechanical response from experimental data remains a fundamental challenge due to non-stationarity behaviour with dynamic regime changes that are weakly handled by standard regression-based approaches. Unlike conventional approaches that predict only peak strength, this study introduces Long Short-Term Memory (LSTM) model used to forecast flexural strength curve of hybrid Macro Basalt Fibre (MBF) reinforced Ultra High-Performance Concrete (UHPC). Using LSTM with dual parallel global-local branches fused with multi-head cross-attention, this framework robustly forecast complete flexural curve across unseen experimental data. In addition, the current work developed comparative computational optimization within LSTM core gates through isolating and evaluating the impact of these gates on the forecasting performance. The results indicate that modifications of some LSTM architecture cells exceed the performance of the baseline LSTM cell regarding error rate and training duration. Where, the variations result in decreased forecasting errors with 57.7% and 30.7% for MSE, MAE, respectively, for LSTM with No Forget Gate (NFG) compared to baseline LSTM. Beyond the studied scheme, these findings create a principle for learning regime-dependent physical behaviour from limited experimental data, offering transferable computational strategy for modelling complex material responses.