I developed a variational autoencoder (VAE)-based model to learn and prioritise promising combinations of chemical elements for synthesis and functional performance. By training on historical chemical data, the model captures underlying patterns and suggests novel phase fields with high potential, reducing experimental trial-and-error. The implementation leverages modern Python ML frameworks (e.g., PyTorch, TensorFlow, Keras) and is fully open-sourced to enable reproducibility and further development. 👉 View code on GitHub
G. Han, A. Vasylenko et. al, Science 383, 6684 (2024) Built a Bayesian optimisation-based framework (PhaseBO) to efficiently navegate complex compositional spaces. PhaseBO leverages Expected Improvement acquisition and Thompson Sampling for batch selection, combined with uncertainty reduction strategies to balance exploration and exploitation.
Compared to conventional grid-searches, PhaseBO enables fine-grained coverage of the compositional space and achieved a 50% increase in the probability of discovering new stable compositions. The implementation is modular and open-source, designed to be extensible for other scientific discovery tasks, such as multi-optimisation or design of experiments.
👉 View code on GitHub
Devised PIGEN - a physics-informed generative framework for crystal structure design based on Denoising Diffusion Probabilistic Models, implemented in PyTorch with multi-GPU support. PIGEN enables generation of crystal structures out-of-distribution, beyond known training data and integrates classifier-free guidance to control generation toward chemically meaningful and stable structures without requiring explicit conditioning labels. The implementation is modular, extensible, and designed for integration with downstream screening or optimisation pipelines.
👉 View code on GitHub