SBOA-GI: A Physics-Informed Enhancement of the Secretary Bird Optimization Algorithm for Gravity Inversion, with an Honest Assessment of Its Limits on Field Data from the Charak Salt-Dome Province, Zagros, Iran

Abstract Metaheuristic optimization is now common for interpreting potential-field geophysical data, yet most published variants are validated only on noise-free synthetic models. This study introduces SBOA-GI, a creatively redesigned, physics-informed variant of the recently proposed Secretary Bird Optimization Algorithm (SBOA; Fu et al. 2024): rather than applying SBOA unmodified, we reconceive four of its five update rules around the ill-posedness, unequal parameter sensitivity, and depth–amplitude trade-off of gravity inversion — to our knowledge the first SBOA variant conceived for a potential-field inverse problem rather than generic benchmarks. The redesign combines Latin-hypercube/opposition-based initialization, self-adaptive differential mutation, sensitivity-weighted Lévy flight, a Lévy–Cauchy stagnation-escape mechanism, physically bounded exploitation with an embedded local Levenberg–Marquardt (LM) polish, and a shrinking-bound restart loop. SBOA-GI, standard SBOA, particle swarm optimization (PSO), and LM were compared on synthetic fault, sphere, and horizontal-cylinder models (0/5/10% Gaussian noise, 30 runs each) and on real Bouguer gravity data (875 stations) from two salt-dome anomalies near Charak, Zagros fold-thrust belt, Iran. At zero noise, SBOA-GI cut parameter RMSE roughly eleven orders of magnitude versus standard SBOA; the gain shrank to negligible at 5% noise and reversed at 10%, showing its refinements mainly speed convergence rather than raise the noise-limited accuracy ceiling. On real data, contrary to expectation, classical LM gave the lowest and most stable misfit (≈ 2.16 and ≈ 1.25 mGal RMS) at a fraction of the metaheuristics' cost, reflecting a smooth misfit surface and the well-known depth–amplitude trade-off. We report this transparently as evidence for the practical limits of metaheuristic enhancement in gravity inversion.