Self-Assigning Spectra with Deep Learning
Assigning molecular spectra is a necessary evil: we obtain a wealth of information about chemistry and physics through tedious and error-prone manual analysis by searching for patterns and assigning spectral features to transitions between quantum mechanical states. I developed a set of recurrent neural network models with reinforcement learning to “teach” a computer how to solve spectra.
Teaching Neural Networks
I created a github repository of Jupyter notebooks and environment that teaches the basics of deep learning: starting from low-level implemtations with
jax, and building up to abstractions with PyTorch and beyond. Experiments can be run locally, or in the cloud with Google Colab.
Probabilistic Molecule Identification
Identifying molecules from unknown mixtures is difficult—I developed a stack of probabilistic deep learning models that help transform experimental spectroscopic information into features that can help uncover the identity of new molecules.
Unsupervised Molecule Discovery in Astrochemistry
Predicting what molecule to look for next is a difficult thing to do when chemical and physical conditions of interstellar environments can potentially be highly complex. In this project, I build a machine learning pipeline for predicting molecular properties by leveraging cheminformatics and data science tools to predict and direct astronomical observations and laboratory experiments.