Structure-Preserving Excess Gibbs Learning for Multicomponent Phase Equilibria: A Falsifiable and Uncertainty-Aware Synthetic Study

Abstract

This fully synthetic study tests a structure-preserving surrogate on NRTL-generated phase equilibrium systems; it contains no experimental molecular validation. A symmetric neural potential represents GE/(nRT), from which activity coefficients, thermal derivatives, binodals, flashes, and tangent-plane-distance diagnostics are derived. Across ten binary training seeds, its median lnγ MAE was 1.16×10−3, better than matched direct regression (2.40×10−3) but not Redlich–Kister (9.65×10−4) or correctly specified refitted NRTL (2.36×10−4). Gibbs–Duhem residuals decayed as O(h2) under grid refinement, confirming diagnostic truncation around an analytic identity. A 41,600-point binary hull gave endpoint MAE 2.31×10−4. On 100 ternary type-I feeds, phase-count agreement with numerical reference labels was 98.0% (exact 95% interval, 93.0–99.8%), with two false negatives after recovery of one infeasible-floor solver failure. A 20-member ensemble showed pooled error ranking (ρ=0.894) but weak in-domain ranking (ρ=0.164); a temperature–distance heuristic was stronger (ρ=0.914), and nominal 90% conformal coverage fell from 88.0% in-domain to 0.0% under shift. Controlled flashes were 2.77× slower than NRTL. A hard thermodynamic structure is therefore valuable within neural modeling, but classical dominance, shift sensitivity, and absent molecular data bound the claim.