Skip to content

ML Researcher & Engineer

Xiaokun (Ken) Zhong

Optimization, Evaluation & Scientific Discovery

I study how machine learning models fail and how to improve them through optimization and systematic evaluation. My research focuses on training dynamics and failure modes in scientific machine learning.

I’m interested in systems that use reliable evaluation and experimental feedback to guide the next experiment and improve their methods. I’m currently working on SciML autoresearch.

Research Engineer in ML Systems & Scientific AI.
Research with Dartmouth and Berkeley collaborators.
BSc Mathematics, Computer Science Track @ HKUST.

CV GitHub Google Scholar LinkedIn Email

Research

Understanding and improving learning systems

Published research · ICML 2026

Understanding when—and why—optimization methods fail

Research question
Why do scientific ML models fail under different training conditions, and which optimization methods help in each regime?
Key finding
Across the SciML models studied, we identify three distinct training regimes. Optimization effectiveness depends on the regime: no single method performs well across all three.
My contribution
As a co-first author, I investigated failure modes, implemented second-order optimization pipelines in PyTorch, and built reproducible Slurm workflows for ablation studies and model behavior analysis.

Unveiling Multi-regime Patterns in SciML: Distinct Failure Modes and Regime-specific Optimization

Y. Wang*, Y. Hu*, X. Zhong* et al.

* Equal contribution

arXiv Code

Ongoing research

SciML Autoresearch

I’m working on automated research for scientific machine learning, extending my interests in optimization and model evaluation toward automated scientific discovery.

Work in progress · Paper not yet public.

News

Recent updates

  1. Our scientific ML paper was accepted to ICML 2026.

  2. Conducted scientific machine learning research with Dartmouth and Berkeley collaborators.

  3. Graduated from HKUST with a BSc in Mathematics, Computer Science Track.

Experience

Research and education

Research Engineer, ML Systems & Scientific AI

Dartmouth College · UC Berkeley collaborators

Feb 2025 - Present

Berkeley, CA

  • Investigated failure modes and training instability in scientific AI systems under stiff PDE and other challenging regimes.
  • Implemented second-order optimization pipelines in PyTorch and scaled reproducible experiments on Slurm-based HPC clusters.

Research Engineer, Medical Imaging AI

HKUST Undergraduate Research Program

Feb 2023 - May 2025

Hong Kong

  • Developed deep learning workflows for MRI reconstruction and 3D volumetric modeling from sparse medical data.
  • Improved Python and C++ data processing pipelines for faster experimentation on large imaging datasets.

BSc in Mathematics, Computer Science Track

Hong Kong University of Science & Technology

Sep 2022 - Jun 2026

Hong Kong

  • Coursework includes differential equations, artificial intelligence, and probability theory.
  • Visiting student in Computer Science at UC Berkeley from Dec 2025 to May 2026.

Additional Projects

Other engineering work

Medical Imaging AI

Deep learning models and data pipelines for MRI reconstruction and 3D volumetric modeling from sparse medical data.

Secure Network Tunneling & Transport Optimization

Cross-region networking system using TLS-based transport and system-level tuning for reliability and latency analysis.

Autonomous Robotics Vision System

Real-time computer vision and embedded control work for target acquisition in RoboMaster robotics.

Contact

Research collaborations and ML roles

I’m interested in research collaborations and ML roles in optimization, model evaluation, and automated scientific discovery.

xzhongal@berkeley.edu GitHub LinkedIn CV