Gage DeZoort

AI × Physics Postdoc @ Princeton

Machine learning researcher and physicist specializing in graph neural networks, graph transformers, and scalable deep learning for scientific data. I build principled architectures with an emphasis on optimization stability, hyperparameter transfer, and robust inference, with applications in collider physics and astronomy.

Resume CV Email
Portrait of Gage DeZoort

Current Work

Coupled Particle-Edge Networks

I am extending hyperparameter transfer parameterizations from standard GNNs to models with attention and edge features — Coupled Particle-Edge Networks (CPENs). The aim is the same as the published transfer work: tune a small proxy, then scale.

Jet classification

Tagging collider jets with CPEN architectures that inherit learning rates from smaller, cheaper-to-tune models.

Stellar streams

Identifying stellar streams in astronomical surveys with the same transfer rules.

Plenary talk, HAMLET-PHYSICS 2026

Publications

QM9 results showing learning-rate transfer across GNN sizes and improving train loss and validation MAE as models scale

Hyperparameter Transfer in Graph Neural Networks 2026

G. DeZoort and B. Hanin, arXiv:2607.05017 (2026)

A recipe for scaling up GNNs using a hyperparameter transfer parameterization.
Test accuracy versus depth on Cora and stochastic block models, comparing vanilla GNNs with residual aggregation and residual connections

Principles for Initialization and Architecture Selection in Graph Neural Networks 2025

G. DeZoort and B. Hanin, SIAM Journal on Mathematics of Data Science 7, 1 (2025)

Training deep GNNs through careful initialization, mitigating oversmoothing, and overcoming correlation collapse.
Detector cross-section showing particle tracks, tracker hits, and calorimeter layers

Graph Neural Networks at the Large Hadron Collider 2023

G. DeZoort et al., Nature Reviews Physics 5, 281 (2023)

Surveying state of the art GNNs applied to a broad range of particle physics tasks.

Heavy pseudoscalar Higgs search (A → Zh → ℓℓττ) 2025

CMS Collaboration, Journal of High Energy Physics 10, 074 (2025)

Search for a heavy pseudoscalar Higgs boson in the full CMS Run 2 dataset.

Charged particle tracking via interaction networks 2021

G. DeZoort et al., Computing and Software for Big Science 5, 26 (2021)

Charged particle tracking via edge-classifying GNNs.

Selected tutorials

Gage DeZoort teaching a practical machine learning crash course at UC Irvine

Crash Course in Practical ML 2025

A three-day UCI crash course (100+ attendees) on practical ML, taught as a School of Physical Sciences Visiting Fellow.
Students following along during the Princeton graph neural networks workshop

Graph Neural Networks for Your Research 2022–2025

A recurring Princeton Research Computing workshop on applying graph neural networks to scientific research.

Creative Side Quests

Guitar has been a lifelong hobby. I keep a public profile of my playing at @gagedezoort.music.

Gage DeZoort performing live on guitar
Gage DeZoort playing a progressive metal set
Gage DeZoort playing bass under stage lights

Contact