Abstract: A brief introduction to an increasingly popular machine learning technique, SVM (support vector machine) is presented for solving nonlinear classification and regression problems. Properties of SVM are discussed together with potentials of applying SVM to numerical weather forecast. SVM is a novel learning method that has solid theoretical basis and requires only small amount of sample. It does not rely on probability measures and Law of Large Numbers, hence is different from many other statistical methods. In essence, SVM smartly evades the traditional inference process from induction to deduction. Instead, it employs transductive inference from training sample to predicting sample, which greatly simplifies classification and regression problems. The decision function of SVM is only determined by a few support vectors. The complexity of computation relates to the number of support vectors rather than the dimension of the sample space. Thus, to some degree, SVM avoids the "curse of dimensionallty".