About me

Welcome! I am an assistant professor in the Department of Industrial Engineering and Management Sciences (IEMS) and the Department of Mechanical Engineering (ME) at Northwestern University, where I lead the Integrative AI (IAI) Lab. I obtained my Ph.D. from the Department of Industrial and Operations Engineering at the University of Michigan. The research in our lab is driven by the need for novel AI and statistical methodologies addressing scientific and engineering challenges across diverse domains, including Digital Twins and smart manufacturing. We are also interested in investigating the theoretical underpinnings of these methods. In particular, our current research in personalized and generative data analytics explores computational techniques to integrate knowledge from multiple sources.

Featured papers

A few topics of my research are introduced below.

DiffATS: Diffusion in Aligned Tensor Space Jinhua Lyu, Tianmin Yu, Brian Kim, Lizhuo Zhou, Chanwook Park, Naichen Shi, 2026. Link, Code.

Training diffusion generative models directly on high-resolution spatiotemporal data is expensive. DiffATS addresses this challenge by moving diffusion from the original data space to a compact aligned tensor factor space.

The core idea is to represent each sample by data-adaptive low-rank tensor primitives. These primitives are extracted directly from the data through Tucker decomposition and then stabilized by Orthogonal Procrustes alignment.

Generated Results

We show that the proposed primitive maps are homeomorphisms: the representation inherits the statistical patterns from the original data while compressing them by up to 210 x.

Scalable Mean-Field Variational Inference via Preconditioned Primal-Dual Optimization Jinhua Lyu, Tianmin Yu, Ying Ma, Naichen Shi, 2026. Link.

The scale and complexity of modern Bayesian statistics are growing. While Mean Field Variational Inference (MFVI) offers an efficient approach to probabilistic inference, it struggles to handle hierachical models where the number of latent variables grows with the number of data points. We developed a mini-batch version of MFVI and introduced a primal-dual algorithm, PD-VI, along with its block-preconditioned extension, P²D-VI. These methods utilize an augmented Lagrangian formulation to jointly update global and local parameters while adapting to heterogeneous loss geometries.

Mouse

On a spatial transcriptomics domain detection task, P²D-VI is capable of segmenting transcriptomic profiles across over 150,000 locations in under 6 seconds.

Here is a more comprehensive list of publications. You can also check my Google scholar profile.

Naichen Shi

Featured papers

Recent news