Human-Constrained AI-Assisted Identification of Defect-Topology Relations for Mechanical Degradation in Porous Solids: A Discrete-Element Study

Abstract

Porosity measures how much material is removed but not how defects interrupt load-bearing paths. We develop a human-constrained, AI-assisted workflow to identify compact topology–degradation relations in a controlled bonded-particle system with 10,040 specimens spanning five pure families and mixtures. Investigators defined the descriptor grammar, admissible information, physical prior, model size, and approval rules; language models proposed and critiqued descriptor concepts; authors curated and implemented the 127-channel bench; deterministic TRAIN-only code selected terms and fitted the ordinary-least-squares relations. On the reused, non-independent reporting benchmark, the six-descriptor representations produced higher scores than the evaluated three-dimensional convolutional network under the matched protocol. Applied without reselection or refitting to 900 porous specimens from three post-freeze packing realizations, the relations achieved prospective numerical R2=0.516 for strength-decay rate and 0.735 for modulus-decay rate, with greater family heterogeneity for strength. Across families, modulus-decay rate was predicted more consistently than strength-decay rate. Together with the exact contrast k=kE+Δk, this predictive asymmetry motivates, but does not establish, a two-stage failure hypothesis. The exponential form remains a compact empirical parameterization rather than an established physical law. All evidence is numerical and confined to the tested YADE/JCFpm setting; validation across numerical configurations and experimental systems remains necessary.