Machine Learning for Shallow-Foundation Settlement Prediction: A Critical Systematic Review of Evidence, Validation and Engineering Readiness

Shallow-foundation settlement prediction is nonlinear, data-limited and dependent on soil, foundation and loading context. Machine learning accuracy is difficult to interpret when evidence sources, independent units and evaluation designs differ. This systematic critical review searched Scopus and Web of Science through 20 August 2026. Revision-stage reconciliation of the original search records yielded 96 primary reports spanning forward, inverse and explicitly settlement-derived targets, with eight adjacent-outcome papers analysed separately. The primary reports mapped to 64 analysis/provenance codes; 44 reports occupied 12 recurrent codes, and the historical SPT lineage supported 32 reports. These codes are not counts of statistically independent datasets. Fifty-three documented matched comparison settings were identified across 47 reports; rankings remained specific to the common target, evaluation cases and metric. Domain-specific appraisal distinguished missing reporting from demonstrated methodological weaknesses. Separate-source validation and calibrated predictive uncertainty remained limited, while explainability did not establish causality or external validity. A review-derived R0–R5 framework separates generator fidelity, internal and grouped empirical evaluation, external physical evaluation, prospective field testing and monitored use. The evidence supports candidate decision support within evaluated domains, not routine substitution for engineering design.