Yale University · New Haven, CT

Steven Zhang

Molecular biophysics, applied mathematics, and machine learning.

GitHub LinkedIn

San Francisco, 2026

About

I’m a rising junior at Yale University studying Molecular Biophysics & Biochemistry and Applied Mathematics. I’m interested in computational chemistry, computational biology, and machine learning.

During summer 2026, I worked as a software engineering intern at OnyxPoint Global Management and conducted research through the NYU Simons Center for Computational Physical Chemistry REU in the Hocky Group. At Yale, I’m part of the Kueh Lab, which studies the molecular circuits that shape immune-cell decisions.

Research

Computational chemistry

Learning structural fingerprints for crystal discovery

Built and evaluated a scalable pipeline for classifying simulated colloidal crystals, matching known structures, and flagging unfamiliar structural families for further investigation.

Systems immunology

Modeling immune-cell decisions

Using stochastic simulation and mathematical models to reason about how immune cells integrate signals, proliferate, differentiate, and form memory.

Selected projects

Reinforcement learning

Learning to fight.
Learning to survive.

Tank Duel

I built a tank-combat environment and trained a neural agent to navigate, aim, dodge, and fire around ricocheting bullets. Demonstrations initialize the policy; PPO refines it, with physics-based safety masks to reduce self-hits.

The green tank is the latest selected agent, facing an earlier trained agent in a maze. This is a recorded match, played at normal speed (1×).

74.7%: Wins against hard shooters
224 of 300 test games
0 / 1,600: Observed self-hits
Across the final test suites

Fixed-map benchmarks against scripted and frozen neural opponents; not a human win rate or a guarantee of safety.

Explore the project

Green: trained agent · Red: earlier trained agent · Silent, looping replay

Starting on opposite sides of a wall, both agents move toward an opening and exchange fire. Each fires three shots; green wins without a self-hit. Use the video controls to pause or replay.

Computer vision · Transfer learning

Teaching a foundation model to recognize blood cells.

BioMedCLIP × BloodMNIST

I compared zero-shot predictions, a classifier trained on frozen image embeddings, and partial fine-tuning of BioMedCLIP. Updating the final visual transformer block, normalization, projection, and classifier improved performance on eight blood-cell classes.

Same test set. More useful representations.

Frozen linear probe88.89%

Partial fine-tuning96.99%

Test accuracy · 0–100% · 3,421 images

Saved results from one completed run. The zero-shot baseline was evaluated on validation data, so it is not included in this test-set comparison.

Read the experiment

Eight labeled BloodMNIST training images: basophil, eosinophil, erythroblast, immature granulocyte, lymphocyte, monocyte, neutrophil, and platelet. — One training example from each of the eight classes. These are dataset labels, not model predictions.

Teaching

Can a Computer Taste?Splash at Yale · 2026

Splash at Yale · 2026

Can a Computer Taste?

Machine Learning Through Food

Students turn their own taste ratings into a dataset, then train models to predict enjoyment—from linear regression to a small neural network.

60 minutes
Grades 6–12
Pair programming

Class notebook & materials

Inside the class E5492

01
Taste & rateBuild a dataset from real responses.
02
Guess & trainCompare your weights with linear regression.
03
Reserve & testTry wasabi peas and check unseen answers.
04
Build & compareCode a small neural network in Colab.

More projects

Molecular Toxicity GNNGraph attention network on Tox21 molecular data GitHub OpenADMET ChallengeExperiments in molecular property prediction GitHub YourFoodieConversational restaurant discovery application GitHub

Off the bench

I keep notes on ideas, papers, and questions that I want to return to.