I’m Sylvester Zhang — a PhD candidate in Chemistry at McGill University. I use computational chemistry and machine learning to figure out how materials work and how we can make them work better. When I'm out of the lab, you’ll find me outdoors camping, sailing, or biking. I love combining problem-solving with fun, whether it's in the lab or on an adventure.
Some of my research projects include:
- Methane Dehydroaromatization on Gallium Nitride
- High-throughput simulations to understand the intermediates of methane reaction dehydrogenation on Gallium Nitride.
-
Methane is cheap and everywhere. Making it useful is hard and expensive. Gallium nitride (a semiconductor in your phone charger) can rip methane apart and turn it into useful benzene - but how?
- How well do small drugs dissolve in plastics
-
Simulating with GROMACS how drug molecules dissolve into plastic amphoteric nanoparticles.
-
Collaboration to improve drug delivery systems (work embargoed)
- Gallium Imide for Methane Activation
-
Sometimes weird molecular structures can efficiently rip apart methane.
-
How does this happen? What can we do to make these molecules even better?
- SO3-equivariant Neural Networks for Transition State Prediction
-
Real life has rotational symmetries (ie the SO(3) group). Neural networks that are equivariant to rotational transformations can predict chemical reaction pathways.
-
Most chemistry neural networks use graph neural networks (with or without symmetry). At transition states, the assumptions that bonds can be properly described as graph edges breaks down.
-
Why waste compute describing physically questionable assumptions? Just learn directly on position at transition states to save time and energy!
-
Project currently underway
- LLMs in Chemistry
-
How well large language models can help in chemistry?
-
LLMs are powerful, but severely limited by what they have seen.
-
Supervised Finetuning, token level entropy matching and retrieval augmented generation can all significantly improve performance
- Bayesian optimization of ligands for proteins
-
Graph genetic algorithm + tabinmoto kernels quickly generate plausible ligands for various proteins
-
trade wars Bayesian optimization is good and easy to win!
Some of my personal projects:
-
Automated Job Search Tool
- Scrapes job listings, matches them to provided interests, and autowrites resumes and cover letters using LLMs.
- Vector search, RAG, and agentic decision-making with LangChain and DSPy.
-
Reinforcement Learning for the Royal Game of Ur
- Train deep Q-learning and SARSA agents to play one of the oldest board games ever.
