Answers are grounded in this paper and cite their sources.
Sequential excavation of twin tunnels creates a staged settlement problem because the second drive acts on ground already affected by the first. We developed VF, a source informed dual expert machine learning surrogate that learns first stage settlement (S1) and incremental second stage settlement (ΔS = S2 − S1) from separately trained Gaussian and Boltzmann derived representations and reconstructs cumulative settlement S2 exactly. In nested five fold evaluation of 52 staged HSsmall numerical scenarios, VF predicted ΔS with R 2 = 0.839 and RMSE = 0.2315 cm, nearly identical to the Gaussian expert (0.2324 cm) and lower than the Boltzmann expert (0.2578 cm). Additional conventional baselines showed that a raw design ExtraTrees model achieved lower RMSE for S1 and S2 (0.9492 and 1.0592 cm) than VF (1.0761 and 1.1856 cm), whereas VF retained the lowest ΔS RMSE among the added baseline set (next lowest 0.2379 cm for gradient boosting). Thus, the evidence supports response dependent representation use rather than universal ensemble dominance. Blocked family testing exposed poor extrapolation for Isfahan S1 and S2 when complete response families were absent from training. VF is supported as a leakage controlled numerical surrogate within the audited HSsmall domain; independent field validation remains required before physical predictive use.
No takes yet. Share an insight, caveat, or question.
Rostami et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: