Comparative Study of Machine Learning Methods for Nonlinearity Prediction in Nanopositioning

Nanopositioning is of great importance in precision engineering and nanotechnology, but the positioning accuracy is severely deteriorated by the ubiquitous nonlinearities such as the hysteresis nonlinearity of the actuator as well as the nonlinear relation between input and output of the closed-loop control system. This paper focuses on the nonlinearity prediction in nanopositioning by using different machine learning methods. The prediction capability of three emergent machine learning models including the Gaussian process (GP), the neural network (NN) and the Koopman operator (KO) are comparatively studied by experiments on a self-made nanopositioner. The indexes of the nonlinearity prediction performance to be evaluated are the prediction accuracy and the prediction efficiency. Experimental results are summarized and discussed, which could provide a reliable selection guidance of the nonlinearity prediction methods, and thus facilitating the feedforward compensation in the precision motion control of nanopositioners.