TabPFN Extensions for Interpretable Geotechnical Modelling

The embedding-based similarity analysis is compelling: the cosine-similarity heatmap clearly separates Clay and Sand samples without explicit soil-type supervision at the embedding level, and critically, reveals out-of-distribution uncertainty for Test sample No. 8 that the probability surface fails to capture. The SHAP analysis convincingly recovers physically interpretable two-regime structures—index properties dominate consolidation parameters while cross-parameter influence governs strength predictions—matching established geotechnical relationships like the inverse dependence of preconsolidation pressure on water content. The native posterior distributions show physically reasonable parameter-specific uncertainty (broad for $C_{\mathrm{v}}$, narrow for $C_{\mathrm{c}}$).

The classification task uses a synthetic dataset where shear-wave velocity $V_{\mathrm{s}}$ is deterministically derived from N-value via prescribed empirical formulae, making the two classes perfectly separable and the task artificially trivial—accuracy of 1.00 is expected, not impressive. The sample size is extremely small (32 total samples, 16 train/16 test), raising questions about statistical significance. The iterative imputation procedure lacks a convergence criterion—the 10-iteration limit is arbitrary—and assumes well-calibrated posteriors that could propagate errors. The paper acknowledges but does not resolve the critical limitation that TabPFN's decision boundaries fail in regions lacking training data coverage, as shown in the low-N, low-$V_{\mathrm{s}}$ region where the model incorrectly predicts Sand despite no training examples.

Furthermore, the claim that "formal calibration analysis is deferred to future work" is concerning given that the iterative procedure's validity depends on calibration. The regression benchmark only uses 20 test samples, and no comparison with established uncertainty quantification methods (e.g., conformal prediction, Bayesian neural networks) is provided.

The evidence supports the claim that TabPFN provides interpretable insights, but the experimental design limits generalizability. The embedding analysis (Figure 3) and SHAP visualizations (Figures 6-7) effectively demonstrate interpretability. However, the comparison with existing approaches in Table 3 is qualitative and self-serving—'conventional ML' is vaguely defined, and TabPFN's lack of retraining is compared against methods that would require it, without acknowledging that foundation models amortize training costs across pretraining. The companion study [12] is cited for accuracy benchmarks, but this paper provides no quantitative comparison of interpretability or uncertainty quality against hierarchical Bayesian models which are the established geotechnical standard. The claim that SHAP provides "complementary, model-agnostic perspective" compared to hierarchical Bayesian models is accurate, but the paper does not demonstrate whether this added granularity improves decision-making.

Reproducibility is severely limited. No code, data, or trained model checkpoints are provided. The synthetic classification dataset generation—deriving $V_{\mathrm{s}}$ from N-values using Japanese railway seismic design standard formulae—is described but not reproducible without the specific standard reference [11]. Random seeds for the train/test split are not reported. Critical hyperparameters for the iterative imputation (number of iterations $K=10$) were "set without a formal convergence criterion." The SHAP analysis uses the permutation explainer from tabpfn-extensions, but permutation count and other configuration details are omitted. The benchmark dataset BM/AirportSoilProperties/2/2025 is referenced but not accessible. For independent reproduction, researchers would need: (1) the exact synthetic dataset or generation script, (2) tabpfn-extensions version and SHAP configuration, (3) iterative imputation stopping criteria, and (4) random seeds.