Wesley Reinhart1,Arindam Debnath1
The Pennsylvania State University1
Wesley Reinhart1,Arindam Debnath1
The Pennsylvania State University1
Refractory High-Entropy Alloys (HEAs) are a promising class of materials for ultra-high-temperature applications including energy generation from gas turbines. In addition to having exceptional mechanical properties at elevated temperatures, these materials can be highly tailored to individual applications by selection of the constituent elements. The relationship between elemental composition and function is challenging to understand and even harder to predict because it is nonlinear, high-dimensional, and results from physical phenomena at many scales. As a result, machine learning is an attractive tool for the empirical design of these materials. While conventional materials design has utilized predictive models to rapidly test hypothesized material compositions in a search for improved ones, more recent generative modeling approaches provide for the possibility of “inverse modeling.”<br/><br/>We have recently developed deep-learning-based generative models including variants on the Generative Adversarial Network (GAN) architecture to perform inverse modeling of refractory HEAs with tailored properties. Generative modeling offers an attractive solution to materials design problems due to its ability to approximate the inverse function directly (i.e., properties to composition) without the need to search the design space. However, we have also faced challenges in training our models on sparse and uncertain experimental data gathered from literature. Here we compare the design of HEAs using these generative models (the “inverse” paradigm) and surrogate regression models (the “forward” paradigm). We discuss the lessons learned from our preliminary work and strategies we are developing to measure the effectiveness of each approach in designing new HEAs for ultra-high-temperature applications.