Fig. 3

Flowchart of the AMOSA algorithm for 1-D TEM inversion. The process starts from reading the observed TEM data, followed by the initialization of temperature, definition of objective functions (model fitting and model-constraint), generation of initial models, and archive initialization. Perturbation sampling based on the temperature-dependent Quasi-Cauchy distribution is used to produce new models. For each new model, 1-D TEM forward modeling is performed and both objectives are evaluated. The dominance relationship between the new and current models determines the acceptance strategy: deterministic acceptance when the new model dominates, probabilistic acceptance when dominated, and conditional acceptance when non-dominated but dominating at least one archived model. The archive is updated accordingly. Iterations continue with temperature reduction until the fitting error is below a predefined threshold. The final inversion result is obtained by averaging the top three Pareto-optimal models with the smallest data misfit.