<p>With the application and development of multibody theory, research on dynamic simulation and optimization design of multibody systems has attracted increasing attention and interest. The dynamics equations of complex multibody systems are typically modeled as high-dimensional and strongly nonlinear differential-algebraic equations, which makes dynamics optimization particularly challenging. Solving them using numerical integration methods is computationally expensive. In recent years, the machine learning involving physical information has been widely applied to solve partial differential equations with promising results. In this paper, an efficient multibody system machine learning (MSML) method for dynamics optimization is proposed. Based on MSML, high-accuracy predictions for dynamic simulation and sensitivity analysis are achieved, enabling efficient optimal design of multibody system dynamics. The dynamic optimization examples of the slider-crank mechanism and the space net system validate the promising engineering application prospects of the proposed method.</p>