Dynamic Mask-Guided Residual Routing for Concrete Strength Evolution

Learning concrete strength evolution from literature-derived data is challenging because complete multi-age trajectories and isolated single-age records often coexist across heterogeneous sources, while random record-level splits can overestimate generalization. We propose Dynamic Mask-guided Residual Routing (DMRR), a compact neural framework for heterogeneous concrete-strength modeling. DMRR aligns different material descriptors into a shared mixture-age representation that supports both trajectory inputs and record-level observations. It introduces dynamic feature masking, which uses a differentiable concrete relaxation to learn sample-adaptive feature gates for robust and interpretable feature selection. It further employs depth-wise attention routing to aggregate hidden representations from multiple network depths, replacing fixed residual reuse with input-dependent routing. A soft monotonicity regularizer discourages physically implausible strength degradation while tolerating experimental noise. Evaluated on a three-dataset concrete temporal suite, DMRR is tested under group-by-mixture validation, long-age extrapolation, cross-dataset transfer, and ablation settings. Results demonstrate improved prediction and interpretable feature selection patterns for age-dependent concrete strength evolution.