Conditional Generative Modeling for Inverse Design of High-Entropy Alloys with Tailored Hardness

We have recently proposed the use of conditional Generative Adversarial Networks (cGANs) to perform inverse design of refractory High-Entropy Alloys (HEAs) with tailored properties. The cGAN architecture offers an attractive solution to materials design problems due to its generality. In effect, the design rules for any material can be learned by this model given sufficient examples on which to train. With only a few hundred observed HEA compositions reported in literature, our model was able to capture important trends in the data and reproduce realistic-looking alloy compositions. However, our preliminary model struggled to extrapolate to novel combinations of properties, and in our initial study we were not able to validate the designs produced by the model experimentally.<br/>We now consider the hardness of the designed alloys as a case study in the predictive accuracy of the model. This property is often reported and can be readily measured experimentally, yet accurate models which consider processing conditions of HEAs are not broadly available. Thus, we consider properties that influence the hardness such as phase, temperature, and processing scheme as inputs to the model and evaluate their effect on the overall performance.<br/>While training on real data is desired, the required information may not be available in literature for all the HEAs. We also address this sparsity in the dataset through data imputation by surrogate models. We compare the performance of the model trained on the smaller, experimental dataset versus the larger but uncertain data obtained through imputation. The trained generator is deployed to generate candidate compositions that will be experimentally manufactured and tested for hardness. The validation by experimental characterization will result in a new training dataset for our model and enable the continuous generation of more tailored novel alloys, with each iteration increasing the probability of arriving at the target mechanical properties.