Abstract
<title>Abstract</title> <p>Vanadium oxides are a compositionally rich family of functional transition-metal oxides with diverse electronic, optical, thermochromic, electrochemical, and structural behaviors. Their multiple accessible oxidation states and complex phase space make them attractive targets for computational materials discovery, but make systematic first-principles screening computationally demanding. We propose an inverse-design workflow that couples a voxel-based variational autoencoder (VAE) with a formation-energy-constrained Wasserstein generative adversarial network (WGAN) to generate physically plausible V--O crystal candidates while mitigating adversarial training instability. Crystal structures are encoded as element-resolved 3D voxel grids, compressed into a continuous latent space, and sampled by a WGAN whose generator is biased toward low formation energy via an exponential penalty predicted by a frozen convolutional neural network (CNN) regressor. Using 10,981 density functional theory (DFT)-relaxed structures for training and validation, the framework generates 6,592 candidate V--O crystals with high generative quality (validity 81.5%, uniqueness 96.6%, novelty 99.5%). First-principles screening identifies 669 stable candidates (10.2%), 365 metastable candidates (5.5%), and a further 94 fall below the Materials Project--referenced convex hull and are reported as candidates for further validation. A composition-matched DFT control confirms that the formation-energy constraint raises the first-principles stable fraction from 15.9% to 25.4%. Phonon calculations on representative candidates, including one below the hull, support lattice-dynamical stability, with only minor soft modes attributable to finite-size effects or known phase transitions. The proposed pipeline reduces the DFT screening burden and provides a scalable route to functional oxide compositions for thermochromic, electronic, electrochemical, and adaptive thermal-material applications.</p>