Ben Thorne

EXPERIENCE

Lead Software Engineer (prev. Senior MLE) 2023–present

Atomic Industries

Geometry & physics-based optimization of plastic injection molds
  • Led the end-to-end design and implementation of an internal tool for automated design of injection molds, combining advanced physics, computational geometry, and optimization.
  • Architected core data structures, simulation pipelines, optimization algorithms and parallel implementation in Julia. This included:
    • A Monte Carlo solver for elliptic PDEs, enabling scalable thermal simulation in highly detailed 3D geometries.
    • A highly optimized geometry package for 3D spatial queries and acceleration structures, used to evaluate physical constraints and manufacturability in real time.
    • A distributed evolutionary algorithm framework tailored to the discrete, constrained nature of the design space.
    • Various optimization algorithms for subproblems including mixed-integer linear programming, gradient descent, and reinforcement learning.
  • Acted as technical lead for a small team, setting direction, running sprint planning, and reviewing code.
  • Collaborated closely with manufacturing engineers to validate simulation outputs against physical behavior, and worked with product leadership to support customer PoCs and Series A fundraising with design outputs.

Machine Learning Engineer 2023

NERSC, Lawrence Berkeley National Laboratory

FAIR Universe project (NeurIPS’24) [PDF]
  • Coordinated the FAIR Universe project: a cross-institutional collaboration of 10+ scientists and engineers from Berkeley Lab, University of Washington, Paris-Saclay, and ChaLearn to build a ML challenge platform for scientific computing.
  • Contributed to the design of uncertainty-aware ML benchmarks and scoring metrics for high-energy physics datasets, shaping the first public challenge launched at NeurIPS 2024.
  • Worked closely with engineers at NERSC to support the deployment of containerized ML payloads on the Perlmutter supercomputer, using Kubernetes and Rancher for orchestration.

Postdoctoral Researcher 2019 - 2022

University of California, Davis

Generative modeling of dust maps using VAEs in pytorch [PDF] [GitHub]
  • Curated novel training dataset of 1,000 images of the interstellar medium from public data.
  • Designed and trained convolutional variational autoencoder in PyTorch.
  • Applied trained model to Bayesian inverse problems: data imputation, denoising, inference.
Differentiable physical models for cosmology with CUDA and auto-diff [PDF]
  • Developed automatically-differentiable foreground model extension to CMBLensing.jl.
  • Implemented GPU-acceleration with CUDA.jl.
  • Developed sparse approximations & preconditioners to speed up log likelihood evaluation by \(\sim 100\times\).
  • Distributed Bayesian inference pipeline across 10’s of A100 GPU nodes on Perlmutter.
  • Used to analyze \(10^6\)-pixel 3-channel images from BICEP/Keck-South Pole Telescope joint analysis.
Python Sky Model (pysm) simulation package [PDF] [GitHub] [Project Page]
  • Original author of the Python Sky Model package, pysm, for simulating microwave sky maps. This package has become the de facto standard for simulating cosmic microwave background surveys, with over 300 citations.
  • Uses numba and MPI for execution in high-performance computing environments.
  • Frequently runs across 100’s of nodes at NERSC in large-scale simulation campaigns for the Simons Observatory and Stage-IV experiments.
  • Due to widespread community reliance on the package, since 2020 it has been maintained and developed by the Pan-Experiment Galactic Science Group, an academic consortium.
Other duties
  • Co-organized the weekly cosmology seminar from January 2020 to December 2022.
  • Co-supervised PhD students on various projects in machine learning and physics.
  • Regularly delivered seminars and conference talks.

EDUCATION

PhD in Astrophysics 2015 - 2019

University of Oxford, Princeton University & Kavli IPMU

During my PhD I worked on various aspects of Cosmic Microwave Background data analysis, supervised by Professor Jo Dunkley. I was a Kavli IPMU - Oxford fellow, meaning that my time was split between the University of Tokyo’s Kavli Institute for the Physics and Mathematics of the Universe, and the University of Oxford. I also spent two years at Princeton University as a visiting student.

Performance forecasts for the Simons Observatory [PDF]
  • Wrote maximum likelihood estimation algorithm for pixelized satellite data using Numba.
  • Performed Monte Carlo simulations to forecast constraints on primordial gravitational waves from SO data.
Data analysis for the Atacama Cosmology Telescope [PDF]
  • Computed Bayesian priors for 2018 likelihood analysis by calculating and fitting power spectra of public sky maps.
Novel observables of SU(2)-axion inflation [PDF]
  • Developed analytical and numerical predictions for chiral primordial gravitational wave signals in parity-violating cosmologies.
  • Modified C code, CLASS, to compute parity-violating CMB signatures, and quantified the sensitivies of the upcoming laser interferometer and CMB satellites LISA and LiteBIRD to the derived signals.
  • Resulted in a high-impact paper (100+ citations).

M.S. & B.S. in Physics 2011 - 2015

University of Oxford, New College

I did my undergraduate and master’s studies at New College, Oxford as an academic scholar. In my Master’s year I specialized in theoretical physics and astrophysics. My thesis was supervised by Dr Adrianne Slyz and Professor Julien Devriendt and involved the analysis of galaxy formation simulations.

Photometric decomposition of barred and double-barred galaxies [PDF]
  • Collected dataset of galaxy images from the Sloan Digital Sky Survey.
  • Decomposed galaxy images into components using least-squares fitting, constraining bar and spiral structure.

SKILLS

pytorch, jax, scipy, dask, mpi4py, numba, numpy CUDA.jl, StaticArrays.jl, Distributed.jl, Flux.jl, Zygote.jl : MPI, CLASS Docker, Nomad, Git, DVC, AWS [S3, EC2], Quarto, Javascript, Slurm, MPI