Abel Lawrence Peirson
alpv95 at alumni dot stanford dot edu
[News | Xp | Research | Projects | Misc. | Blog]

I am a recent Ph.D graduate in Physics at Stanford University, where I research high energy astrophysics, astrostatistics and applications of machine learning as a NASA FINESST fellow and Stanford data science scholar. In September 2023, I will be joining Citadel GQS in Miami as a quantitative researcher.

I am interested in scientific applications of machine learning and statistics that incorporate domain knowledge.

One of my ongoing projects has been increasing the science output of NASA's flagship X-ray polarimetry mission IXPE using simulation based inference techniques. I am equally interested in the science made possible by these improvements, for example, how the geometry of relativistic jets affects their polarized emission, and statistical techniques to properly interpret observations. See my research highlights below for an overview.

Previously, I received my bachelors and masters (MPhys) in Physics from the University of Oxford, Christ Church, where I specialised in theoretical and particle physics.



Work/Teaching Experience

  • Google Research -- Brain Team, Student Researcher
    June - September 2022
    Research into practical optimization methods for deep learning.
    Working with Rohan Anil, Ehsan Amid, and Manfred Warmuth. [Webpage]
  • G-Research, Quantitative research intern
    June - September 2021
    Forecasting capital markets and identifying trading strategies. [Webpage]

  • Stanford PHYS113: Computational Physics, Teaching Assistant and Lecturer
    January - April 2021
    Prepared and delivered lectures on optimization and statistical inference. [Slides 1][Slides 2]
  • Stanford PHYS100: Observational Astrophysics, Teaching Assistant
    April - June 2019
    Optical telescopes, practical observation and data analysis.

Open Source Software

  • MLCommons Algorithmic Efficiency: Benchmark and competition for neural network training.
  • ixpeobssim: Simulation and analysis framework for the Imaging X-ray Polarimetry Explorer.
  • MulensModel: Fast gravitational microlensing simulation/fitting package.
  • SSCpol: Polarized relativistic jet simulation in C with Python wrapper.
  • Dank Learning: 'Show and Tell' image captioning for meme generation.


Presently, I am developing second order optimization methods for and understanding the training dynamics of deep neural networks at Google Brain, further improving IXPE's spatial and energy resolutions with simulation-based inference, and interpreting IXPE's ongoing scientific discoveries. Representative publications are highlighted.

Benchmarking Neural Network Training Algorithms
George E Dahl, Frank Schneider, Zachary Nado, et al. (incl. A.L. Peirson)
Arxiv, 2023.

First large scale benchmark for neural network optimization algorithms.
Call to arms for community-wide benchmarking competition.

X-Ray Polarization of BL Lacertae in Outburst
A.L. Peirson, Michela Negro, Ioannis Liodakis, et al. (IXPE collaboration)
ApJ Letters, 2023.

First measurement of X-ray polarization in blazar BL Lacertae.
Measuring polarization in the transition region between synchrotron and inverse Compton emission allows for testing different jet composition hypotheses.

Fishy: Layerwise Fisher Approximation for Higher-order Neural Network Optimization
A.L. Peirson, Ehsan Amid, Yatong Chen, Vladimir Feinberg, Manfred K Warmuth, Rohan Anil
Neurips HITY workshop, 2022.

A local approximation of the Fisher information matrix at each layer for natural gradient descent training of deep neural networks.
Very cool initial work taking advantage of layerwise matching loss functions.

The X-ray Polarization View of Mrk~421 in an Average Flux State as Observed by the Imaging X-ray Polarimetry Explorer
Laura Di Gesu et al. (IXPE collaboration, incl. A.L. Peirson)
ApJ Letters, 2022.

First X-ray polarization measurements of Mrk 421.
The extremely high and constant polarization fraction compared to the optical is exciting -- I'm working on the subsequent observations of this source where we found hints of EVPA rotation.

Polarized Blazar X-rays imply particle acceleration in shocks
I. Liodakis et al. (IXPE collaboration, incl. A.L. Peirson)
Nature, 2022.

Very exciting observations of X-ray polarization in Mrk 501.
I did some of the data analysis for this one and I am designing its potential Nature cover with DALL-E2!

Polarized X-rays Constrain The Disk-Jet Geometry in a Black Hole X-ray Binary
H. Krawczynski, et al. (IXPE collaboration, incl. A.L.Peirson)
Science, 2022.

IXPE observes Cyg X-1, finding a higher than expected polarization -- revealing the structure of its accretion disk and jet.

Testing High-Energy Emission Models for Blazars with X-ray Polarimetry
A.L. Peirson, I. Liodakis, R.W.Romani
The Astrophysical Journal, 2022

Here we show how a well placed IXPE observation could potentially measure both synchrotron and SSC polarization from blazars, probing our current understanding of blazar jets and testing hadronic emission scenarios.

Neural network analysis of X-ray polarimeter data
A.L. Peirson
Springer Nature, 2022

Invited chapter for the Springer Nature textbook "Handbook of X-ray and Gamma-ray Astrophysics", editors: Cosimo Bambi and Andrea Santangelo.

An in depth summary of many of the simulation based inference techniques we have developed for X-ray polarimetry.

A Deep Ensemble Approach to X-ray polarimetry
A.L.Peirson, R.W.Romani
Fourth Workshop on Machine Learning and the Physical Sciences (NeurIPS), 2021

Setting the state of the art signal recovery in X-ray Polarimetry. The code from this and my previous works will eventually be used as part of IXPE's official data analysis pipeline.

We model the uncertainties on heteroskedastic photoelectron track angle predictions using deep ensembles of CNNs and derive the appropriate maximum likelihood estimator for polarization parameter reconstruction. Our work will reduce telescope exposure times by up to 40% for a given signal-to-noise ratio.

New Tests of Millilensing in the Blazar PKS 1413+135
A.L. Peirson, I. Liodakis, A.C.S. Readhead et al.
The Astrophysical Journal, 2021

We fit a complex binary lens magnification model to multiwavelength blazar lightcurves using Bayesian nested sampling (MultiNest). To speed up the algorithm, I used convex optimization to marginalize over the (constrained) linear parameters in real time, reducing the number of free parameters to be sampled. Our empirical results tentatively suggest that the blazar could be lensed.

Towards Optimal Signal Extraction for Imaging X-ray Polarimetry
A.L.Peirson, R.W.Romani
The Astrophysical Journal, 2021

We show our polarization estimators using deep ensemble quantified uncertainties are optimize the signal-to-noise ratio for realistic observations. Using a CNN based classifier we can exclude, with high confidence, events converting outside of the main detector volume. We apply our new technique to a selection of astrophysical spectra, including complex extreme examples.

The Relativistic Jet Orientation and Host Galaxy of the Peculiar Blazar PKS 1413+135
A. C. S. Readhead, ... A. L. Peirson et al.
The Astrophysical Journal, 2021

PKS 1413+135 is one of the most peculiar blazars known. We provide observational evidence suggesting that the blazar could be gravitationally microlensed.

Deep Ensemble Analysis for X-ray Polarimetry
A.L.Peirson, R.W.Romani, H.L.Marshall, J.F.Steiner, L.Baldini
Nuclear Instruments and Methods in Physics Research A , 2020

Our first work using deep ensembles and heteroskedastic uncertainty quantification to improve IXPE's sensitivity by more than 30%.

In order to measure polarization using a gas pixel detector, a distribution of photoelectron angles must be extracted from noisy images of their individual tracks on a hexagonal grid. We use a ResNet-18 architecture to predict the initial photoelectron angle and its uncertainty for each track image, as well as the absorption point and energy. By combining predictions from an ensemble of networks in a weighted maximum likelihood analysis, we estimate the polarization fraction and EVPA of the X-ray source.

The Polarization Behavior of Relativistic Synchrotron Self-Compton Jets
A.L.Peirson, R.W.Romani
The Astrophysical Journal, 2019

Exploring the expected synchrotron self-Compton polarization from blazar jets. A continuation of our multizone geometric jet model.

For low-synchrotron peaked blazars, the X-ray emission will be dominated by synchrotron self-Compton. Understanding what polarization levels IXPE is likely to see in this case is important for distinguishing between hadronic and leptonic blazar emission models. Importantly we find that a rise in synchrotron polarization fraction at high energies is guaranteed by basic relativity considerations.

Prospects for Detecting X-Ray Polarization in Blazar Jets
I.Liodakis, A.L.Peirson, R.W.Romani
The Astrophysical Journal, 2019

A study in the detectability of X-ray polarization in blazars for IXPE.

Using the models developed in the adjacent relativistic jet papers and optical polarization observations by RoboPol, we are able to make predictions of the average expected X-ray polarization fraction. This work helped the IXPE team choose appropriate first year observing targets.

The X-ray Polarization Probe Mission Concept
K.Jahoda, H.Krawczynski, F.Kislat, ... L.Peirson et al.
Decadal Survey on Astronomy and Astrophysics , 2020

A white paper for the upcoming Decadal survey introducing XPP, a second generation X-ray polarimeter that will cover 0.2 - 60keV energies (IXPE covers 1-10keV).

Hopefully we will demonstrate the potential of imaging X-ray polarimeters with IXPE.

The Polarization Behavior of Relativistic Synchrotron Jets
A.L.Peirson, R.W.Romani
The Astrophysical Journal, 2018

Using simple helical geometry and relativistic aberration effects to explain blazar polarization angle rotations.

Our follow up paper for the synchrotron self-Compton emission is above.


Unbiased sample of projects outside of my usual research.

Dank Learning: Generating Memes Using Deep Neural Networks
A.L.Peirson V, E.M.Tolunay
Accepted to Advances in Intelligent Systems and Computing, 2018

An image captioning system that can take any image and turn it into a first generation meme. Uses pre trained Inception CNN followed by an LSTM with temperature selection for language generation. Trained on a dataset of over 300,000 image caption pairs I scraped from memegenerator.net.

What started as a fun class project ended up being featured on Techcrunch and AINews. Inspired, I worked together with Dylan Freedman to create the free iOS app Dank Learning so anyone can use the 'AI' to generate memes. I discuss the work in the NVIDIA AI podcast, and its implications for the future of art and jewellery at the Pratt Institute.

Planet Plotter
Making and displaying unique exoplanet inspired art.

We bought a 1982 HP7470a plotter printer and we make it plot starfields around exoplanets. Check out our gallery of all the exoplanets we know of (we can plot other astronomical objects too... The image on the left is M51: the whirlpool galaxy).


We made a lightweight implementation in javascript of one of our favourite games: Achtung, die Kurve!. We have also made a 3d version in Unity: Achtung, 3die Kurve!, not available online yet. I hope to build an RL agent in collaboration with Taylor Howell that solves Achtung, die Kurve!

  • Ironman 70.3, Santa Cruz 2022. [S: 35m, B: 3hr05, R: 2hr02, T: 5hr55]
  • San Jose Buddhist Judo Tournament, Team Champions 2018.
  • Oxford Summer Eights, Head of the river 2017.
  • Native Spanish speaker.

Inspired by