Accurate estimation of the maximum horizontal stress (SH) is essential for wellbore stability analysis, reservoir development, and subsurface energy applications. Conventional breakout-based methods often rely on elastic assumptions, leading to substantial inaccuracies in formations exhibiting inelastic deformation. This study presents an integrated experimentalnumericalmachine learning framework for SH estimation that explicitly accounts for inelatic rock behavior. Elastoplastic parameters for four lithologies (shale, two limestones, and sandstone) were derived from single- and multi-stage pseudo-triaxial compression tests. A finite element model incorporating the DruckerPrager failure criterion was developed and validated against laboratory data. The validated model was then used to generate a dataset of 2188 synthetic cases linking in situ stress conditions to breakout width and depth. Three transparent Group Method of Data Handling (GMDH) correlations were trained to estimate SH from breakout geometry and rock properties. The best-performing model achieved R = 0.98 and RMSE = 5.89 MPa. Independent validation using modified thick-wall cylinder experiments demonstrated strong agreement between measured, numerically simulated, and machine learningpredicted SH values. In contrast, elastic-based approaches produced significantly overestimated failure zones and unrealistic stress predictions. The proposed framework provides a computationally efficient, field-applicable tool for stress estimation that incorporates inelastic deformation. This approach improves the reliability of breakout-based stress analysis in both brittle and ductile formations and offers practical implementation potential for petroleum, geothermal, and subsurface engineering applications.