Research collection
Building models, then testing what they learn
My chemistry AI work asks two practical questions: can a model learn useful chemical representations, and can it generate molecular structures directly? The collection starts with those projects, then traces the generative-model and scientific-computing work around them.
Start with OLMO learns chemistry01
Chemistry AI
The core projects: language models trained on chemical data and generative models that work directly with molecular graphs.
Language models for chemistry
Teaching OLMo-7B chemistry
Continued pre-training of OLMo-7B on the ChEMBL SMILES training split, followed by candid MoleculeNet classification and regression results with explicit evidence limits.
Read the OLMO project notesMolecular graph generation
MolGAN in PyTorch
A from-scratch implementation of MolGAN with a graph generator, R-GCN discriminator, reward network, WGAN training, and reinforcement learning on molecular graphs.
Explore the MolGAN guide02
Generative methods
This experiment is not chemistry research. It tests an adjacent question: how diffusion can generate structured sequences in a compressed latent space.
Generative AI methods
Latent diffusion for language generation
A poem generator that compresses text with BART and a Perceiver sampler, then trains a smaller SANA diffusion transformer to denoise the latent representation.
See the diffusion architecture03
Scientific computing
These astronomy projects are a separate research lane. They connect the portfolio through scientific data, molecular spectroscopy, and open-source analysis tools.
Spectroscopy and molecular composition
JWST atmospheric chemistry
An analysis of WASP-39b light curves and transmission spectra, including atmospheric retrieval for water, methane, carbon dioxide, and sulfur dioxide.
Explore the JWST analysisScientific computing
Setting up HEASOFT
Practical setup notes for the X-ray astronomy toolkit, including NICER calibration data and the environment needed to process observational files.
Read the HEASOFT notesOpen-source research software
GSoC with Stingray and OpenAstronomy
A proposal for an interactive X-ray binary database that brings light curves, power spectra, energy spectra, and hardness-intensity diagrams into one research workflow.
Read the GSoC proposalDiscuss the research
Questions about a method or interested in collaborating? Email me. You can also visit the author page for more context about my work.