Luke Thompson

PhD Candidate - Computer Science & Applied Mathematics


2024 -
I began my PhD in computer science at the University of Sydney. I work on AI for science with a focus on molecular dynamics and stochastic models. I am also a casual academic teaching third year and masters students in classical machine learning and Bayesian methods.
2024 -
I lead a project to update several legacy VisualBasic data pipelines to a Rust/Postgres backend, achieving a >95% reduction in data processing time. I also maintain Kumon's main geographic data dashboard which guides business decisions.
2023 - 2024
First Class Honours at the University of Sydney. I built the world's best model for mutagenicity prediction from chemical structure using a graph transformer neural network. It was published in an American Chemical Society journal here.
WAM: 86.6 (Thesis, Transcript)
2021 - 2023
Bachelor of Science: Pharmacology and Medical Science.
Talks & Presentations
Discussing AmesFormer at the Australian federal chemical regulator: AICIS (slides)
Discussing AmesFormer at the ACTRA Annual Scientific Meeting in Canberra, Aug 2024, Awarded Student Prize (slides)
My final poster presented as a component of my Honours, Feb 2024
Projects
QuickSig is a Jax library for computing path signatures, log signatures, solving rough differential equations (RDEs), and more. It is currently the most complete library for rough paths, and I am currently working on including two Rust algorithms for computing the Baker-Campbell-Hausdorff series to allow more efficient computation of Lie series for RDE tasks.
Cold War: Iron Curtain is a popular mod for the video game Hearts of Iron IV. I have run a team of ~30 volunteer developers since 2017, recently achieving 600 000 unique downloads in 2023. We're currently one of the most popular cold war strategy games worldwide.
Eigen2 is a Python library for calculating eigenvector-eigenvalue pairs. I implemented >10 different algorithms, including Cholesky decomposition, Givens rotation QR decompositions, and the Tao eigenvector-from-eigenvalue method, among others.
FastDAQ is the world's fastest open-source NASDAQ/ASX ITCH file and datastream parser (excepting FPGA implementations). Built in Rust.
Writings & Selected Notebooks
Publications
Chemical Research in Toxicology - ACS
Luke Thompson, Josiah Evans, Slade Matthews
Other Things I've Done