Benjamin Thorne

Lead Engineer at Atomic Industries, San Francisco

I build optimization and machine-learning systems grounded in physics. At Atomic I work on software that starts from the geometry of a part and designs the injection mold that makes it. We tackle problems at the intersection of geometry, physics, and machine learning, which have no prior industrial scale solutions.

Previously, I was a cosmologist. I worked on modelling the microwave sky at Oxford, Princeton, Tokyo, Davis, and Berkeley.

Email LinkedIn GitHub Scholar CV (PDF)

Current work

A two-plate injection mold, cut away to show the cooling channels and the ejector pins. Scroll the page to open the mold.
Automated tooling designAtomic Industries · 2023–

Most plastic parts are injection molded, and the molds that make them are still mostly designed by hand. At Atomic we are changing that.

I lead the team that writes our mold design software. It takes a customer's part and works out the rest of the tool: where the mold splits, how the cavity is blocked out, and where the cooling and ejection go.

We cover computational geometry, high fidelity physics simulation on GPU, optimization, and machine learning.

Earlier work

Line drawing of a sky patch. A regular grid is bent around a lens mass, and arrows show the deflection of each point.LENSUNLENSED
Gravitational lensing remaps the unlensed sky. The arrows are the deflection field.
Differentiable cosmologyUC Davis · Berkeley · 2020–2022Paper
  • A differentiable foreground extension to CMBLensing.jl, accelerated with CUDA.jl.
  • Sparse approximations and preconditioners cut the cost of a log-likelihood evaluation by roughly 100×.
  • This made inference across tens of A100 nodes practical.
Line drawing of a particle detector seen along the beam. Tracks curve out from the center. A tau jet and a muon are marked, and a dashed arrow shows the missing energy.TAU JETMUONMISSING ECALORIMETER
An H → ττ candidate along the beam: a tau jet, a muon, and missing energy.
FAIR UniverseBerkeley Lab · NeurIPS 2024PaperProject
  • A ten-plus person collaboration across Berkeley Lab, the University of Washington, Paris-Saclay, and ChaLearn.
  • Built uncertainty-aware machine-learning benchmarks for high-energy physics.
  • Served on the Perlmutter supercomputer.
Line drawing of three maps of the whole sky, stacked. The front map shows dust emission along the Galactic plane.353 GHz100 GHz30 GHzGALACTIC PLANE
Simulated maps at 30, 100 and 353 GHz. Thermal dust dominates the highest band.
Python Sky ModelOpen source · 2017–PaperGitHub
  • The de facto standard simulation package for the microwave sky, with 300+ citations.
  • Used by the Simons Observatory, CMB-S4, and the Atacama Cosmology Telescope.
  • Now maintained by the Pan-Experiment Galactic Science group.

Background

2023–now
Lead Engineer, Atomic Industries, San Francisco · previously Senior MLE
2023
Machine Learning Engineer, Lawrence Berkeley National Laboratory (NERSC)
2019–2022
Postdoctoral Researcher, University of California, Davis
2015–2019
PhD in Astrophysics, University of Oxford, with Princeton University and Kavli IPMU
2011–2015
M.S. & B.S. in Physics, University of Oxford, New College · first-class honors

Writing

All publications →