Alzheimer's disease (AD) is a neurodegenerative disorder with limited treatment options. While transcranial magnetic stimulation (TMS) has demonstrated therapeutic potential, its underlying mechanisms and personalized applications are not yet fully explored. This study integrates deep learn-ing with neurodynamics, employing a pretrained resting-state AD self-supervised model to constrain dynamic train-ing and simulate TMS-induced neural dynamics in AD pa-tients. The results reveal regional heterogeneity, showing enhanced excitability in stimulated regions, in contrast to reduced global neural dynamics. We identified significant differences in brain responses between AD patients and healthy controls, which provide critical theoretical support for developing TMS treatment strategies. These insights ad-vance the understanding of AD pathophysiology and high-light the potential of TMS interventions, which may lead to more effective therapeutic approaches.