Accurate description of the free energy landscape (FES) is the basis for understanding complex molecular systems, and for further realizing molecular design, manufacture and industrialization. Major challenges include multiple metastable states, which usually are separated by high potential barriers and are not linearly separable, and may exist at multiple levels of time and spatial scales. Consequently FES is not suitable for analytical analysis and brute force simulation. To address these challenges, many enhanced sampling methods have been developed. However, utility of them usually involves many empirical choices, which hinders research advancement, and also makes error control very unimportant. Although variational calculus has been widely applied and achieved great success in physics, engineering and statistics, its application in complex molecular systems has just begun with the development of neural networks. This brief review is to summarize the background, major developments, current limitations, and prospects of applying variation in this field. It is hoped to facilitate the AI algorithm development for complex molecular systems in general, and to promote the further methodological development in this line of research in particular.