Machine Learning-Driven Discovery of Stable Vanadium Oxide Compositions for Aerospace Applications

The demanding environments of aerospace systems require materials with exceptional properties, yet discovering thermodynamically stable, synthesizable candidates remains a critical challenge for next-generation electronics and protective coatings. We introduce an inverse-design pipeline that couples a voxel-aware variational autoencoder with aWasserstein GAN whose loss incorporates formation-energy constraints, enabling direct generation of viable vanadium-oxide (V–O) candidates from a continuous latent space. The framework encodes atomic positions and lattice parameters on 3D grids, then decodes GAN samples back to real-space structures for ab-initio vetting. Across 1,622 generated compositions, we achieve a remarkable 20% stability rate under strict thermodynamic criteria (negative formation energy and ≤ 300 meV/atom distance to the convex hull), with an additional 15% meeting a metastability threshold (≤ 500 meV/atom). Phonon dispersion calculations on selected stable compositions confirm dynamic stability, with minor imaginary modes attributable to finitesize effects rather than fundamental structural instabilities. The approach yields physically consistent structures with superior generative stability versus conventional GANs, revealing V–O compositions with potential for radiation-tolerant, low-power aerospace systems, thereby demonstrating a transformative computational pathway for developing advanced materials for space-grade electronics, thermal management systems, and radiation-shielding applications.