Data-Driven Delta Machine Learning Models for Improved Extrapolation

Abstract

This work demonstrates a method of testing a machine learning model’s extrapolation accuracy, a capability that is significant to efficiently aid with discovery of improved alloys and presents the application of a pure data-driven method to make steps to reduce this extrapolation error. By using linear models to capture general trends in the data and then the subsequent application of more complex machine learning methods, extrapolation capabilities can be reduced. Being purely data-driven, this type of model can be coupled with other Delta-Machine Learning techniques such as those that utilize physics-domain knowledge, and coupled with active learning methods, with better extrapolation capabilities reducing the number of iterations needed to outperform existing alloys.