James Liu

I build machines
that help us ask
better questions
about nature.

Undergraduate researcher in machine learning for experimental science, at the University of Waterloo.

move to explore

01

About

I work on the decision layer of experimental science: which experiment to run next, and how much to trust the model that suggested it.

Management Engineering (AI option) at the University of Waterloo; currently an ML research intern at the National Research Council of Canada.

02

Research

i

Choosing the next experiment

Constrained multi-objective Bayesian optimization for closed-loop experimental campaigns.

BoTorch · qLogNEHVI · closed-loop design
ii

Physics-informed surrogates

Mechanistic priors with Gaussian-process residual correction, fitted to a handful of runs.

Gaussian processes · hybrid models · uncertainty
iii

Generalization past the training regime

What a learned procedure retains when the problem outgrows its training set.

graph networks · OOD · algorithmic reasoning
03

Experience

  1. May 2026 — Aug 2026 Mississauga, ON

    National Research Council Canada · Clean Energy Innovation

    ML Research Intern

    Physics-informed Gaussian-process surrogates for battery-leaching prediction, plus a constrained Bayesian-optimization loop that cut projected experiments by 83%.

  2. Sept 2025 — Dec 2025 Toronto, ON

    Loblaw Companies Limited

    Data Science Intern

    Probabilistic demand and transport forecasting — 200+ national lanes at 3% WAPE.

  3. Apr 2025 — Sept 2025 London, UK

    University College London

    Undergraduate Research Assistant

    A noise-aware Bayesian optimization pipeline that raised formulation feasibility from 38% to 82%.

  4. Jan 2025 — Apr 2025 Waterloo, ON

    Atlantic Industries Limited

    Process and Data Engineering Intern

    Statistical process control and a real-time Power BI dashboard, cutting production scrap by 20%.

04

Projects

GraphScale

Neural algorithms that generalize to larger problems

Learned graph algorithms tested far outside their training size, where trajectory supervision gained 6.4 points at 256 nodes.

Python · PyTorch · NetworkX · graph neural networks Report  →

05

Publications

  1. Fast-tracking complex formulation development with multi-objective Bayesian optimisation

    H. Liu, A. Gucic, J. H. Liu, M. T. Cook, D. Shorthouse

    Journal of Controlled Release · 2026

    doi:10.1016/j.jconrel.2026.115171

06

Contact

Always glad to hear about a good problem — research collaborations, internships, or a question you think is being asked the wrong way.

j58liu@uwaterloo.ca