Samuel Tovey

I'm a

About

I am a computational physicist from Perth, Western Australia, passionate about machine learning, data science, and Python software development. The next big breakthroughs in science and society will come from the intelligent, innovative, and responsible use of big data. I am looking to apply my physics, mathematics, and data science background to contribute to the next generation of technology.

Data Scientist & Computational Physicist.

I use advanced physics, mathematics, and data science topics to understand big data better. Looking to apply my skills to consulting and result-driven research.

  • Birthday: 18 September 1996
  • Website: www.machinelearningmd.com
  • Phone: +49 151 2026 9164
  • City: Stuttgart, DE
  • Age: 27
  • Degree: Ph.D.
  • Email: tovey.samuel@gmail.com
  • Employment: Available

I was born in Perth, Western Australia, where I completed my Bachelor of Science in physics and electronics engineering. I moved to Stuttgart, Germany, in 2018 to pursue my Masters of Science in Physics. I completed my thesis on developing machine-learned inter-atomic potentials for molecular dynamics simulations. I love to travel, play the guitar, and ride motorbikes in my free time. The photo I have included above is of a Quokka, a marsupial native to my home city. I like to raise awareness of them, not because they are endangered, but because they are the happiest creatures on Earth.

Topic Distribution

I work on a few different topics. Below is my best attempt to quantify the distribution. The 100 % at the end is no math mistake. I write all of my work into user-friendly software packages to ensure it can be deployed in other peoples research straight away.

Theoretical Machine Learning 40%
Multi-Agent Reinforcement Learning 30%
Reservoir Computing and Quantum Machine Learning 10%
Statistical Physics 10%
Atomistic Simulation10%
Software Development100%

Resume

Education

Ph.D. in Theoretical/Computational Physics

2020 - 2024

The University of Stuttgart, Stuttgart, Baden-Württemburg

My doctoral studies focus on the intersection between physics and machine learning. While I participate in many projects, my main research goals are split into two parts. Most of my time is spent developing a theoretical framework for neural networks built on tools in statistical physics. The second part of my research focuses on using reinforcement learning to control microscopic particles to develop technologies for micro-robotics.

Masters of Science in Theoretical/Computational Physics

2018 - 2020

The University of Stuttgart, Stuttgart, Baden-Württemburg

I completed my master's thesis in the development of machine-learned inter-atomic potentials for use in molecular dynamics simulations. This work culminated in a recent publication on an inter-atomic potential for molten sodium chloride.

Bachelor of Science double major in Physics and Electronics Engineering.

2014 - 2017

The University of Western Australia, Perth, Western Australia

During my bachelor's degree, I majored in physics and electronics engineering with a brief specialisation in printed circuit board design and quantum theory.

Projects

MDSuite

SQL, HDF5, Molecular Dynamics

MDSuite is a parallelised, GPU-enabled, memory-safe post-processing package for analysing molecular dynamics simulations.

SwarmRL

Reinforcement Learning, HPC

SwarmRL is a ground-up reinforcement learning and simulation environment for studying intelligent micro-robots.

ZnVis

3D Visualization, Open3D

ZnVis is a visualisation engine built on Open3D. It allows for rendering and visualisation of meshes using a simple pythonic interface.

ZnNL

Theoretical machine learning

ZnNL is a Python package built for performing theoretical machine learning studies. Utilizing Flax, jax, and neural-tangents, ZnNL allows users to take advantage of the research I perform into statistical models of neural networks.

Professional Experience

Research Assistant

2019 - Present

The University of Stuttgart, Stuttgart, Baden-Württemburg

  • Develop a theory to describe neural network training and performance.
  • Utilize multi-agent reinforcement learning to train micro-robots in physics-based simulations.
  • Mentor bachelor's and master's students through their thesis work.
  • Teach classes on fundamentals of computing and software, computational physics, and statistical physics.

Electronics Engineer

2017 - 2018

Brainsystems Electronics Engineering, Perth, Western Australia

  • Design schematics and layouts for printed circuit boards with a specialisation in wireless technology such as BLE and LoRa.
  • Perform physical debugging and testing of printed circuit boards.
  • Perform SPICE simulations of novel schematic design.
  • Work with project tenders to meet customised client needs.

Duty Manager

2014 - 2018

Belmont Oasis Leisure Center/Bayswater Waves Leisure Center, Perth, Western Australia

  • Oversee lifeguards and centre staff members on duty
  • Manage events and classes at the centre.
  • Handle customer complaints.
  • Provide first aid to facility users and manage emergency procedures.
  • Handle dangerous chemicals for the maintenance of pools.

Lifeguard

2014 - 2018

Belmont Oasis Leisure Center/Booragoon Leisure Center/Bayswater Waves Leisure Center, Perth, Western Australia

  • Supervise pools and recreation facilities.
  • Manage day-to-day events.
  • Provide first aid to patrons.

Crew trainer/Staff member

2011 - 2014

McDonalds Corporation, Perth, Western Australia

  • Supervise and train staff members on shift.
  • General maintenance of equipment.
  • Cook and deliver food.
  • Front counter service.

First Author / Equal Contribution Publications

Generating Minimal Training Sets for Machine Learned Potentials

https://arxiv.org/abs/2309.03840

Environmental effects on emergent strategy in micro-scale multi-agent reinforcement learning

https://arxiv.org/abs/2307.00994

Towards a phenomenological understanding of neural networks: data

https://iopscience.iop.org/article/10.1088/2632-2153/acf099

Generating Minimal Training Sets for Machine Learned Potentials

https://arxiv.org/abs/2309.03840

MDSuite: comprehensive post-processing tool for particle simulations

https://link.springer.com/article/10.1186/s13321-023-00687-y

Efficient Data Selection Methods for the Development of Machine Learned Potentials

https://arxiv.org/abs/2108.01582

DFT Accurate Interatomic Potential for Molten NaCl from Machine Learning

https://pubs.acs.org/doi/abs/10.1021/acs.jpcc.0c08870

Invited Talks

Towards a phenomenological understanding of neural networks

ML4Jets2023, DESY, Hamburg, Germany

Multi-agent reinforcement learning for micro-robots with SwarmRL

Max Planck Institute for Intelligent Systems, Stuttgart, Germany

Machine learning for the physical sciences

EPFL Lausanne, Lausanne, Switzerland

Physics meets machine learning

Von Karman Institute for Fluid Dynamics, Brussels, Belgium

XLA and Jax for the acceleration of scientific software

University of Stuttgart, Stuttgart, Germany

MDSuite: High performance computing for the analysis of particle-based simulations

2022 American Physical Society March Meeting, Focus session: Building the Bridge to Exascale, Chicago, USA

Quantum mechanically accurate inter-atomic potentials for molten salts

2021 American Physical Society March Meeting, Online

Supervised Projects

Transformer encoder strategies for multi-agent reinforcement learning with memory

Pre-doctoral research task

This project explores the use of transformer architectures for encoding historial state descriptins in multi-agent reinforcement learning. This project is currently ongoing.

Efficient data representation of protein sequencing

Master's Thesis

This project investigates data representation strategies for use in training large-scale neural network models, including vision transformers and ResNext architectures.

Swarm models for physical reservoir computing

Bachelor's Thesis

This project explores the use of swarm models for physical reservoir computing. The project is currently ongoing.

Understading the role Fisher information in neural network training

Bachelor's Thesis

This work explores the relationship between the Fisher information matrix, the empirical neural tangent kernel, and neural network training. This work is currently ongoing.

Application of Gaussian process regression in the detection of exceptional points

Master's Thesis

In this work, Gaussian process regression was used to identify so-called critical points in Cuprous Oxides. This work was completed in 2023.

Exploring quantum field theory with classical ring polymers

Research project

In this now completed work, a student looked into using ring polymer simulations as a surrogate model for solving problems in quantum field theory.

Exploring the application of random matrix theory in atomistic simulations

Bachelor's Thesis

This work explored applying denoising methods from random matrix theory to atomistic simulations, specifically cross-correlation functions.

Applying deep learning methods to nanopore protein sequencing

Bachelor's Thesis

In this thesis, ResNet architectures were trained to classify protein sequencing data.

Investigating symmetries in neural network representations

Master's Thesis

This work studied the emergence of symmetries after a neural network was trained. Detection of symmetries was also accompanied by the extraction of generators of the symmetry group.

Graph representations for multi-agent reinforcement learning

Master's Thesis

In this ongoing work, graph network architectures are being used to encode the state description of agents in a multi-agent reinforcement learning environment, specifically, micro-robots.

Understadning entropy relations in neural network training

Master's Thesis

This work expanded on the theoretical model I developed in my PhD for understanding neural networks from a statistical physics perspective.

Active learning strategies for machine-learned inter-atomic potentials

Master's Thesis

This thesis explored active learning strategies to develop machine-learned inter-atomic potentials. During this time, we developed a novel approach based on kernel methods for comparing data points in high-dimensional space to identify the most informative regions for exploration.

Environment effects on emergent strategy in micro-robots

Bachelor's Thesis

This work explored the effect of environment on emergent strategy in micro-robots using multi-agent reinforcement learning.

Efficient strategies for data selection for machine-learned potentials

Master's Thesis

In this project, we explored strategies for selecting data from large pools to train machine-learning models on core sets.

Signatures of failure in active learning strategies

Master's Thesis

In this project, various indicators of model failure were assessed to determine their suitability for active learning strategies.

Machine-learned potentials for noble liquid systems

Master's Thesis

In this project, inter-atomic potentials for noble liquid systems were fit and deployed in large-scale simulations.