<p>Automated machine learning (AutoML) methods, particularly AutoGluon, have demonstrated effectiveness for lithology identification by reducing reliance on expert experience through automated parameter determination. However, AutoGluon relies on a fixed set of base learners and a rigid two-layer structure. This paper proposes an AutoML stacking method (ASM) that integrates adaptive base learner selection and adaptive deep multi-layer stacking, providing a novel AutoML framework for lithology identification. Specifically, adaptive base learner selection is achieved through forward or backward search strategies, which are built on the statistical principles of stepwise regression. Adaptive deep multi-layer stacking is automatically optimized via feedforward propagation coupled with validity comparison. To validate the proposed method, experiments were conducted using lithology datasets from the Zagros Basin in the Middle East. The results demonstrated that the forward search strategy surpassed the backward strategy in both accuracy and efficiency. The ASM variant, ASM-B2, leveraging forward-adaptive base learner selection and a two-layer stacking architecture, improved the F1-score by 1.7% relative to AutoGluon and achieved more efficient lithology identification. Further gains were achieved by ASM-BL, which integrates both adaptive base learner selection and adaptive deep multi-layer stacking, resulting in an overall F1-score improvement of approximately 2.04% over AutoGluon.</p>