Predicting the Properties of Materials with Machine Learning
Determining the properties of materials has always been one of the primary goals of research in materials science. Computational models for materials’ property determination are hindered by their high computational cost; it can take weeks (even years) to develop and evaluate a computational model for a single property of a single material. The current approach to this problem is based on what might be called “human learning.” Materials scientists and engineers spend many years gaining experience and build up a set of intuitive rules for “what works when.” Machine learning algorithms offer an alternative to this approach. In the machine learning approach, a large “training set” of materials with known values for the target property is input
for a computer program. This training set is used by the program to gain “experience.” The program identifies patterns in the data and uses these patterns to develop a computational model that fits the data. This computational model can then be used to estimate the properties of materials that are not in the training set. Typically, the program also estimates the error in its prediction.
In my research group, we are currently developing machine-learning algorithms for predicting the biological activity of candidate drug molecules. Our approach is based on support vector regression and the closely related method of Gaussian process regression. The goal of this project is to extend this method to predicting the properties of materials.
The goal of this project is to establish machine-learning methods (e.g., support vector regression) as a tool for predicting the properties of materials. Because the goal is to characterize the computational model, rather than to design materials, the method will be tested for accurately measured properties of well-known materials.
