As a promising way to achieve superior mass sensitivity in RF coils, NMR Microcoil has been frequently implemented in various probehead designs. It has been shown that microcoil-based NMR probes can obtain a mass sensitivity of more than 10 times better than a conventional 5mm Helmholtz coil NMR probe. On the other hand, with the rapid development of modern wireless communications and radar, antennas are becoming more complex, therein having, e.g., more degrees of design freedom, integration and fabrication constraints, and design objectives. Though utilizing different properties of RF waves, some design concepts can be interchanged between NMR RF microcoil and telecommunication antennas. In both designs, full-wave electromagnetic simulation can be very accurate and therefore essential to the design process. However, it is also very time-consuming, which leads to many challenges for coil design, optimization, and sensitivity analysis (SA). Recently, machine-learning-assisted optimization (MLAO) has been widely introduced to accelerate the design process of telecommunication antennas. Machine learning (ML) methods, including Gaussian process regression, support vector machine (SVM), and artificial neural networks (ANNs), have been applied to build surrogate models of antennas to achieve fast response prediction. With the help of these ML methods, various MLAO algorithms have been proposed for different applications. In this study, we first introduce this concept into the design of RF microcoil for NMR application. Through utilizing MLAO, several RF coils design were optimized theoretically for samples with different geometries. Some simple RF microcoils are also manufactured to validate the simulation.