Modeling physical, chemical or biological systems is probably among the great challenges in science. When an expertise of the system of interest is available, a knowledge base model can be designed. It is often founded on a physical analysis and consists of mathematical equations derived to describe the system. However, when the physical mechanisms running in the system are not fully known, machine learning techniques are very valuable if rich and relevant sets of measurements are available. They allow building efficient nonlinear models with a wide range of applications. The architecture and the design will be described in the following steps : (i) feedforward models using neural nets, (ii) classifiers with support vector machines and (iii) semi-physical models for dynamic modeling of processes. Various fundamental contributions, original methodologies and applications to real and complex industrial systems will be presented.