The estimated national cost due to methamphetamine (meth) abuse is 23.4 billion each year. Current law enforcement of dealing with the meth abuse is by 'fire fighting' strategy, that is, to direct the enforcement to the location after the problem occurs. A better strategy is to develop a proactive predictive model based strategy. This study attempts to develop a model to predict the risk of meth production for the seven most severe states in USA. Modern data mining supervised modeling techniques, including Decision Tree, Generalized Logistic Regression and Neural Network are applied to model the risk level of meth at the block group geographical area. The final best model found is a neural network model. The prediction accuracy is found at 87.5%. Fourteen important inputs are identified, which are further classified into three factor groups: culture, socio-economic and geographical factor groups. It is found that the culture factors, which reflects the specific behavioral models of people, are the most important for contributing to the risk of meth, and that the meth production activity is highly associated with crimes.