I work on AI for Science, geometric deep learning, and data-driven mechanical metamaterial design, with a focus on generative models, multimodal reasoning, and human-AI scientific discovery systems.
My research goal is to build general-purpose AI systems that can reason about, generate, and evaluate mechanical metamaterials from multimodal and multi-objective perspectives.
Email: wzhan24 [at] vt [dot] edu
Office: D&DS, Virginia Tech
GitHub: wzhan24
Google Scholar: Scholar Profile
I am a Ph.D. student in Computer Science at Virginia Tech, advised by Prof. Dawei Zhou. My research lies at the intersection of AI for Science and geometric deep learning, with a focus on data-driven and knowledge-guided mechanical metamaterial design.
I develop unified frameworks that integrate 3D topology, density conditions, and mechanical properties for efficient and controllable material generation and evaluation. My recent work explores diffusion models, symbolic reasoning, and human-in-the-loop systems for automated hypothesis generation, structure refinement, and interactive scientific discovery.
We propose UniMate, a unified generative model for mechanical metamaterial design that simultaneously handles 3D topology, density conditions, and mechanical property prediction.
MetamatBench is a comprehensive benchmark for metamaterial discovery, combining multimodal datasets, machine learning models, and a visual-interactive interface for human-AI collaboration.
METASCIENTIST is a human-in-the-loop framework combining large language models, symbolic reasoning, and 3D structure diffusion for hypothesis-driven metamaterial generation and refinement.
MetaRef is a physics-aware refinement framework that improves the geometric regularity of machine-generated metamaterials, including symmetry, periodicity, and connectivity, through a Lattice Sensor and Physics-Aware Refiner.
This Blue Sky paper presents a vision for generative intelligence in metamaterial design, arguing for a transition from costly trial-and-error workflows to autonomous, data-driven exploration and invention.
DuetDA is a decomposed and dynamic data attribution framework that assigns memorization and generalization values to scientific data, using model-state gating to accelerate training and improve out-of-distribution performance.
PHYVER is a physics-grounded material claim verification system that translates free-form material claims into executable multi-fidelity physical evidence, including MLIP-based optimization and DFT computation.