Nitrate (NO₃) leaching is a major issue in sandy soils intensively cropped to potato. Modelling could test the effect of management practices on nitrate leaching, particularly with regard to optimal N application rates. The NO₃ concentration in the soil solution is well known for its local heterogeneity and hence represents a major challenge for modeling. The objective of this 2-year-study was to evaluate machine learning regression methods to simulate seasonal NO₃ concentration dynamics in suction lysimeters in potato plots receiving different N application rates. Four machine learning function approximation methods were compared: multiple linear regressions, multivariate adaptive regression splines, multiple-layer perceptrons, and least squares support vector machines. Input candidates were chosen for known relationships with NO3 concentration. The best regression model was obtained with a 6-inputs least squares support vector machine combining cumulative rainfall, cumulative temperature, day of the year, N fertilisation rate, soil texture, and depth.