Machine Learning | Astrophysics | Data Science

I am currently incorporating deep-learning techniques within probabilistic programming systems so as to better perform statistical inference within complex forward models specified by simulations. Probabilistic programming is a programming-language paradigm in which programs are written such that they correspond exactly to Bayesian posterior densities. These programs can then be evaluated many times to estimate the posterior, with a plethora of different techniques used 'under the hood'. Importance sampling is a simple sampling technique that is guaranteed to converge, but that can be very slow with the wrong proposal distribution. Using pyprob I use simulation inputs/outputs themselves to train deep neural networks to optimize the proposal distributions. As a side effect of this work, I am also familiar with a wide variety of modern statistical techniques as well as modern data science. I also have real-world experience working for Thymia where I was part of a team that developed a machine-learning pipeline to assess the mental health of app users.

As an astrophysicist, I am interested in how non-linear cosmological structure formation can be understood using inspiration from *N*-body simulations.
These simulations are extremely useful, but are too expensive to be run for every cosmological scenario under consideration.
I have worked on 'rescaling' methods to alter the cosmology of an existing simulation by remapping length and time units and modifying the internal structure of haloes.
I also developed an augmented version of the halo model to produce accurate non-linear matter power spectra, which are useful for analysing weak-lensing data.
This code is publicly available HMcode and rapidly provides non-linear spectra at high accuracy.
HMcode is also incorporated within CAMB.

I am currently using deep learning to infer orbital parameters in Kepler systems with strong transit-time variations. These occur in systems with massive planets with relatively close orbits, such that the orbits interact with each other and can be strongly non-Keplerian.

HMcode-2020: Improved modelling of non-linear cosmological power spectra with baryonic feedback

Mead, Brieden, Troester, Heymans

MNRAS, 2021

Spherical collapse, formation hysteresis and the deeply non-linear cosmological power spectrum

Mead

MNRAS, 2017

Mead, Heymans, Lombriser, Peacock, Steele, Winther

MNRAS, 2016

A full list of my academic publications can be found here.

You can email me at alexander.j.mead@googlemail.com

My data science CV can be found here and my academic CV can be found here.