A self-modelling system for material research has been developed based on discriminant analysis, artificial neural networks, classification mapping and genetic algorithms. It provides systemic methodologies for nonlinear multivariate modelling and multi-objective optimizing. It is designed to unveil connotative information from a limited experimental data set and gives qualitative, quantitative and geometry models of the object to be researched. In addition, optimized research schemes can be derived from these models by genetic algorithms and classification mapping. The technique is suitable for subjects that have some original study results but for the following reasons there are difficulties in doing further research. (i) The object researched has too many controlling factors and is too complex to analyse. (ii) The object is controlled by some unexplainable mechanisms and is difficult to analyse. (iii) The mathematical expression has strong nonlinearity and is difficult to resolve strictly.