Attention-based hybrid deep learning framework for modelling the compressive strength of ultra-high-performance geopolymer concrete

Ultra-high-performance geopolymer concrete (UHPGC) offers a promising low-carbon alternative to Portland-cement-based ultra-high-performance concrete, yet its mechanical behaviour is governed by a highly non-linear interplay among mix proportions, activator chemistry and curing age. To support data-driven mix-design optimisation, we curated and released a benchmark dataset containing 427 distinct UHPGC mixes with compressive-strength records at eight curing ages (3-360 days). Leveraging on this, we propose a hybrid deep-learning framework that cascades one-dimensional convolutional neural networks (CNNs) and long short-term memory (LSTM) layers with a Feature Masking Attention (FMA) module. The CNN extracts local compositional patterns, the LSTM captures latent temporal dependencies across ages, and FMA randomly masks input features during training while learning adaptive attention weights, thereby enhancing generalisation and interpretability. Five-fold cross-validation demonstrates that the CNN-LSTM-FMA model achieves an R² of 0.904 ± 0.018, an RMSE of 5.56 MPa and an MAE of 3.38 MPa, surpassing conventional deep-learning baselines and ablated variants by 4-16%. Moreover, an analysis of the model’s attention mechanism, corroborated by SHAP importance analysis, reveals that cement (Cem), coarse aggregate (Cag), and ground-granulated blast-furnace slag (GGBS) are the most influential factors governing strength development. The proposed dataset and model constitute an open, extensible platform for AI-assisted, low-carbon concrete design.