This paper presents a method for automatically generating new designs from a set of existing objects of the same class using machine learning. In this particular work, we use a custom parametric chair design program to produce a large set of chairs that are tested for their physical properties using ergonomic simulations. Design schemata are found from this set of chairs and used to generate new designs by placing constraints on the generating parameters used in the program. The schemata are found by training decision trees on the chair data sets. These are automatically reverse engineered by examining the structure of the trees and creating a schema for each positive leaf. By finding a range of schemata, rather than a single solution, we maintain a diverse design space. This paper also describes how schemata for different properties can be combined to generate new designs that possess all properties required in a design brief. The method is shown to consistently produce viable designs, covering a large range of our design space, and demonstrates a significant time saving over generate and test using the same program and simulations.