Numerical modeling and simulation are central to the activities of Électricité de France (EDF), both to support the ecological transition and to ensure the safety, reliability, and integrity of energy production facilities. In this context, the growing complexity of nonlinear models makes the setup of simulations increasingly difficult for the design engineers, especially for post-buckling studies. Post-buckling analyses require several steps, including a preliminary buckling analysis, the introduction of geometric or material imperfections, the selection of a continuation (arc-length [1, 2]) method [3], and the definition of the arc-length increment list. These choices strongly affect both solver convergence and the computed equilibrium paths, in addition to that they currently rely heavily on user expertise. A novel approach based on reinforcement learning [4, 5] is proposed to automatically adapt the arc-length increment size during the analysis, with the objective of improving the quality of equilibrium path tracking while preserving convergence. Preliminary results obtained on parametrized curves highlight the potential of reinforcement learning for this application. Our work focused specifically on the reward shaping to encapsulate the definition of a good increment size. The objective is for the agent to be able to give increment size in order to capture all the limit points of the equilibrium path (e.g. points with horizontal and vertical tangents). Future work will focus on extending the approach to geometrically nonlinear structural analyses, defining efficient and physically meaningful reward functions. The aim is to transfer the training done on parametric curves (training environment) directly on the response curve of a given quantity of interest (e.g. force-displacement curve) for mechanical analysis problems (test environment).