Machine-learning (ML) models are increasingly proposed for the bearing capacity of foundations on rock masses, but the evidence for design use has not been appraised for rock foundations specifically. This narrative critical review links the mechanics of foundations on Hoek–Brown rock masses to an appraisal of the ML studies that target them. Of 275 unique Scopus and Web of Science records screened by one reviewer, 48 addressed ML for foundations on rock (full texts read for 11); 21 studies are tabulated with their data origin, size and validation. The review is not systematic. Numerical data sets contain 80 to 5,120 runs, whereas field or literature-compiled data sets for rock-socketed piles and shafts contain 96–200 records. Sample counts and inputs recur across studies. One study generated Gaussian-noise records from its full 136-shaft set and then split the combined set 85/15; three studies tested on the same measured shafts without stating overlap with the training database, the latest using qu as an input although the target is Nσ = qu/σci. A three-level validation hierarchy is adapted to rock foundations. An illustrative synthetic experiment, not a bearing-capacity model, shows that row-wise random cross-validation of project-structured data can give R² ≈ 0.996, whereas grouped cross-validation gives 0.68 ± 0.14. An error of ±5 units in the Geological Strength Index changes the Hoek–Brown constant s by factors of 1.74 and 0.57. Within the sources examined, ML serves mainly as a surrogate for numerical solvers; independent generalization and field validity have largely not been demonstrated.
Undrained shear strength (S u ) is often estimated from index properties, but reported machine-learning (ML) accuracies frequently rest on random splits that place rows from the same site in both training and test sets. Using the public SH-CLAY/11/4051 database of Shanghai clay, we evaluate eight regression models (ridge regression, support vector regression, Gaussian process regression, random forest, XGBoost, LightGBM, a perceptron and a perceptron ensemble) for predicting field vane S u from depth, plasticity index, void ratio and effective vertical stress (377 complete rows from 28 sites). Models are assessed by nested leave-one-site-out cross-validation, against log-normal multivariate (ln-MVN) and linear mixed-effects baselines using identical inputs, with site-cluster bootstrap intervals. All ML models outperform the ln-MVN baseline: pooled out-of-site root-mean-square error (RMSE) is 5.15 to 5.34 kPa for the nonlinear models, 5.64 kPa for ridge regression and 6.09 to 6.10 kPa for the ln-MVN baselines (paired differences 0.4 to 0.9 kPa, 95% intervals excluding zero). Differences among the nonlinear models are small; only the ANN ensemble and random forest clearly outperform the mixed-effects baseline, and adding depth improves no model. Split-conformal intervals reach 0.88 to 0.91 coverage at the 90% level, but coverage varies strongly between sites. Effective vertical stress dominates out-of-site permutation importance. Monotonicity constraints did not change accuracy; an offset estimated from five local measurements lowered the remaining error by 0.1 to 0.4 kPa. The work applies existing methods to an existing database under deployment-consistent validation; it has no external validation set.
Geosynthetic-reinforced soil (GRS) walls are designed mainly through strength-based limit-state checks, yet serviceability is governed by wall deformation, which depends on reinforcement stiffness, facing, foundation, water and time. This critical narrative review, which is not a systematic review, asks how well the evidence on GRS-wall deformation, failure, durability, reliability and machine learning (ML) supports engineering use. It synthesises 72 bibliographically verified sources and grades them by how directly they observe real walls (five levels, E1–E5). Two failure databases attribute 98% (171 walls) and 99% (320 walls) of failures to improper design or construction rather than to material defects. Three ML studies of wall deformation use 166 finite-element analyses with the same four inputs, two of them explicitly the same analyses, and report accuracy from random splits, which supports interpolation within one numerical dataset and not prediction of real walls. A six-level validation scale (L0–L5) states what a reported error can support; none of the graded sources reached the levels that test extrapolation or real walls (L2–L5). A synthetic experiment shows, for one function only, that a high random-split score can coexist with negative R2 under hold-out. The review offers a deformation-centred framework and a measurable research roadmap, and states its limits: the search was not logged and grading was done by one author.