Inverse Design of Crystal Compositions via Improved Generative Models

The discovery of new materials is fundamental for technological advancements in healthcare, sustainable energy, and electronics. Traditional exploration methods, relying on chemical intuition, are slow and inefficient for meeting modern demands. Recent progress in deep learning, combined with extensive datasets, has transformed materials discovery. Generative models facilitate the inverse design of materials by predicting properties from existing data. In this study, we developed an improved Generative Adversarial Networks (GAN)-based model to generate new vanadium oxide (V-O) binary compositions. Our model focuses on stability, utilizing screening techniques and a trained formation energy predictor, ALIGNN, to reduce DFT calculations. A comparative analysis with a vanilla GAN demonstrates that our model is significantly more efficient, generating nearly 1.5 times more stable V-O structures. This highlights the potential of advanced generative models in accelerating the discovery of new materials.