Generative Adversarial Networks for Generation of Synthetic High Entropy Alloys

High-performance materials are key tools for several reasons. On the one hand, their use brings obvious progress in the performance of the pieces where they are used in fields such as aeronautics, construction, or biotechnology. On the other hand, high-performance materials also allow more efficient use of energy in industrial processes where the use of such energy becomes intensive with its consequences in terms of environmental and economic sustainability. For these reasons, the emergence of high-performance materials such as high entropy alloys (HEAs) has captured the attention of industry and researchers within the last years. However, the development of these materials requires a large amount of time and money invested in the design, synthesizability evaluation, construction, and characterization of such compounds. The use of artificial intelligence for the design of materials, even in its current infancy status, provides a valuable tool to accelerate the initial phases of materials design and HEAs, where the high number of combinations brings a perfect scenario for the deployment of machine learning techniques. In this work, a generative-based approach is used, namely generative adversarial networks (GANs), to generate synthetic HEAs for highly intensive industrial processes. The architecture model of a GAN involves two neural networks. The first one is a generator model for generating novel composition candidate alloys to form the HEAs. The second one is a discriminator model for classifying the generated samples coming from the generator in real or fake compositions. The discriminator learns from a specific data structure that contains data from real samples to classify the novel generated samples. A GAN extension that conditionally generates the synthetic outputs by the addition of extra inputs was used. This so-called conditional tabular generative adversarial network (CTGAN) was developed to be used with tabular datasets as input. Such data is normally composed of a mix of continuous and discrete columns, making some deep neural network models fail in performing proper modeling for this kind of inputs. In the present approach, the generated realistic synthetic data was based on the conventional parametric design parameters used for HEAs, i.e., atomic size difference, mean atomic radius a, average melting temperature Tm, mixing enthalpy H_mix, mixing entropy S_mix, electronegativity, valence electron concentration (VEC), mean bulk modulus K, and the standard deviation for most of them. As conditioned input data, the chemical composition of the alloys and their phase has been considered. The phase was classified into four classes, namely amorphous, intermetallic, solid solution, and solid solution + intermetallic, which can be used as an indicator for their applicability. The CTGAN provides novel output candidates of HEAs, the expected parameters mentioned above, and the corresponding phase. The generated data is compared with the calculated data and a verification of novel generated compositions is done in open materials databases available in the literature. Finally, a specific data structure for the CTGAN training and results of the performance of this approach is provided, which was developed in the framework of the European project ACHIEF for the discovery of novel materials to be used in industrial processes.