Machine-learning researcher

Chin-Wei Huang

I work on structured machine learning across scales of atomistic modeling, from electronic structure and molecular simulation to proteins and materials. I am a senior researcher at Microsoft Research.

01 / Overview

I study how mathematical and physical structure can make learned models tractable, scalable, and useful.

I develop generative models and architectures that encode geometry and physical constraints. The goal is not only to fit complex scientific data, but to build models whose structure supports the computation the science requires.

I completed my PhD at Université de Montréal and Mila with Aaron Courville. My thesis, Probability flows in deep learning, developed likelihood-based generative models through deterministic and stochastic transformations of probability distributions.

02 / Current work

Learning across scientific scales

Current projects apply structured machine learning at different levels of atomistic description, from electron densities to proteins and crystalline materials.

02

Molecular simulation

Two for One

Diffusion models that are also force fields

Coarse-graining makes molecular simulation less expensive by replacing groups of atoms with a smaller set of variables, but it also requires recovering the effective interactions lost in that reduction.

Two for One trains a diffusion model on equilibrium structures. Its conservative score serves two purposes: it generates independent equilibrium samples and defines a force field for molecular dynamics. The same model reproduces equilibrium distributions, free-energy landscapes, and folding behavior on benchmark proteins.

Two for One: Diffusion Models and Force Fields for Coarse-Grained Molecular Dynamics JCTC, 2023
03

Materials design

MatterGen

Generative modeling for inorganic materials

Materials discovery is an inverse-design problem: rather than evaluating a known crystal, the goal is to generate stable structures with desired properties.

MatterGen is an equivariant diffusion model for periodic crystals. It can generate diverse, stable, and novel materials, or condition generation on chemical, mechanical, electronic, and magnetic properties. Candidate materials from the model were synthesized experimentally, connecting generative design to physical validation.

A generative model for inorganic materials design Nature, 2025

03 / Foundations

Probability flows in deep learning

My PhD studied likelihood-based generative modeling through change of variables. The central question was how the structure of a probability transformation controls the tradeoff between flexibility and scalable computation.

Deterministic transport

Structure makes expressive maps tractable.

Normalizing flows learn invertible maps between a simple reference distribution and a complex data distribution, allowing both sampling and likelihood-based training.

One way to design such maps begins with the Knothe-Rosenblatt rearrangement. Its triangular structure reduces multivariate transport to a sequence of monotone one-dimensional transformations and gives a tractable triangular Jacobian. I used this structure in Neural Autoregressive Flows, replacing affine transformations with expressive monotone neural networks, and proved the first universal approximation theorem of normalizing flows.

Neural Autoregressive FlowsICML, 2018 Oral
Paper

A distinct route comes from Brenier's theorem: the optimal transport map is the gradient of a convex potential. Convex Potential Flows use this structure to obtain invertibility through strong convexity, inversion through convex conjugacy, and scalable likelihood gradients through convex optimization.

Convex Potential FlowsICLR, 2021 Spotlight
Paper

Augmentation is orthogonal to either choice of transport map. By introducing auxiliary variables, an invertible model can define a more flexible marginal distribution while retaining a tractable variational likelihood bound. The resulting position-auxiliary-variable construction is analogous to phase-space dynamics, and I showed that Augmented Normalizing Flows can approximate a Hamiltonian ODE as a universal transport map.

This connection has since been used in molecular modeling. Timewarp treats velocity-like variables as non-physical auxiliaries in a learned time-coarsened dynamics model, while scalable Boltzmann generators combine augmented coupling flows with graph neural networks for equilibrium sampling and free-energy calculations in large materials systems.

Augmented Normalizing Flows arXiv, 2020 Theoretical precursor presented at the ICLR DeepDiffEq Workshop, contributed talk

Stochastic transport

Diffusion extends probability flows to noisy dynamics.

Score-based models formulated through stochastic differential equations emerged as a promising approach to generative modeling, but initially lacked a likelihood-based theoretical treatment. I derived a continuous-time variational framework that unifies diffusion models with hierarchical variational autoencoders and continuous normalizing flows. Within this framework, I proved that minimizing the score-matching loss is equivalent to maximizing a lower bound on the likelihood of the learned reverse SDE. This result applies to a family of reverse processes whose deterministic limit is the probability-flow ODE.

A Variational Perspective on Diffusion-Based Generative Models and Score MatchingNeurIPS, 2021 Spotlight
Paper

We then extended this variational framework to general Riemannian manifolds, proving its equivalence to Riemannian score matching and deriving the associated probability-flow ODE. We also developed scalable methods for estimating Riemannian divergence without the closest-point projections required by earlier approaches. The resulting models apply across spheres, tori, hyperbolic spaces, and rotation groups, including protein and RNA backbone angles.

Riemannian Diffusion ModelsNeurIPS, 2022
Read the doctoral thesis Probability flows in deep learning