Supplementary Research

Unsupervised Representation Learning

This paper presents relevance encoding networks (RENs), a novel probabilistic framework for VAEs, addressing latent dimensionality mismatch without exhaustive hyperparameter tuning. By leveraging automatic relevance determination (ARD) priors in the latent space, RENs directly learn data-specific bottleneck sizes, ensuring high-quality representation and generation without compromising model flexibility.

RENs: Relevance Encoding Networks
arxiv

Machine Learning for Material Science

Collaborative interdisciplinary materials science project with Idaho National Labs, guided by Dr.Tolga Tasdizen. My research was to develop an innovative combinatorial framework of molecular dynamics and machine learning that explores a large chemical-configurational space to evaluate the mechanical properties of multi-component alloys. I initiated the development of the dataset for atomistic simulations.