Abstract by Parker Hamilton
Physics and Astronomy
Simplifying Materials Prediction with Dimensionality Reduction
One challenge of using machine learning to predict materials properties is understanding what exactly an accurate model can tell us about the systems being studied, especially when using neural networks. Using kernel ridge regression and a neural network trained across 10 binary alloy systems, we show that the initial layers of the neural network can be used as a reduction of the models original input vectors with a minimal loss of information. This allows us to both better understand the machine learning model and use these reduced representations for more efficient calculations.