Answers are grounded in this paper and cite their sources.
Axisymmetric roof-collapse free boundaries couple nonlinear rock strength, plastic dissipation and gravitational work. We formulate a variational neural approximation for a specified Hoek–Brown tensile-correction family whose dissipation is defined by closed convex relaxation. A finite-energy Fenchel identity characterizes the fixed-radius minimum, and a scalar minimum-net-power equation selects a unique positive critical radius for 0 < μ ≤ 1. Local estimates establish recovery of the original Hoek–Brown geometry and relate exact continuous residuals to the radius error. The constrained neural representation preserves roof contact, nonnegative slopes and a material-defined flat core while parameterizing the outer boundary and admitting a core-edge jump. Verification uses ten published A/B-law cases and 51 archived GHB evaluations representing 13 normalized problems and 39 distinct trained parameter states; 12 records restore existing states at other dimensional scales. In the historical informed-initialization cohort, the maximum global geometric error is 0.24%, while exact jump errors reach 37.65%. Separate neutral-jump continuations at μ = 0.20 retain exact jump errors of 68.45–73.27%, despite small integrated residuals. Independent P1 Ritz results and a layered example delimit the scope of the neural approximation; accurate global geometry does not imply accurate recovery of the local jump.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: