<p>Lateral double-diffused MOSFETs (LDMOS) are widely used in power management circuits. However, co-optimizing the static and switching properties of the LDMOS still faces challenges. This paper presents a novel deep reinforcement learning (DRL)-based multi-objective optimization framework. We first developed a deep neural network (DNN) trained on technology computer-aided design (TCAD) simulation data to address the nonlinear relationship between device design parameters and electrical characteristics. The DNN surrogate model is used to interact between the agent and the DRL environment. Subsequently, we incorporated the surrogate model into the soft actor-critic (SAC) algorithm, enabling intelligent optimization of device parameters through automated exploration of the design space. The effectiveness and practicality of the proposed DRL-based framework for device optimization were validated through comprehensive TCAD simulations. In addition, a typical switch circuit was established in the mixed-mode simulation in TCAD. It demonstrated that the energy consumption of the switch circuit decreased by 15.6%. The proposed method demonstrates inherent flexibility that facilitates adaptation to various device optimization tasks.</p>