PhD Studentships

This project will aim to develop robust data handling, analysis, and modelling tools for the automation of high-power laser interactions. These tools will enable the real-time optimisation of plasma accelerator performance, through by closed-loop machine learning algorithms. Plasma accelerators have the potential to revolutionise particle accelerator technology due to their extremely high accelerating fields and short pulse durations. Harnessing data-driven algorithms to optimise plasma accelerators will enable them to reach their full potential by controlling the highly non-linear plasma physics at their core.

The software tools developed during this project will include control of the properties of a high-power, ultra-short laser pulse and the plasma target with which it interacts. Real-time analysis of the experimental diagnostics will be used to build machine learning models of the accelerator performance, and then to guide the optimisation of the interaction to produce particle beams ideally suited to their applications.

The developments of this project will be used to enhance the utility of plasma accelerators at the Extreme Photonics Application Centre (EPAC), a newly constructed STFC facility. The generated particle beams will be used for novel scientific and industrial applications such as ultra-fast x-ray spectroscopy and radiobiology, rapid x-ray tomography and gamma-ray imaging.

University

Queen’s University Belfast

Facility

CLF

Diamond Light Source is a world leading synchrotron facility and part of the UK scientific infrastructure. The synchrotron is comparable to an enormous microscope, which uses x-rays to look at the atoms inside molecules and materials. Each day, teams of researchers from Universities around the world bring their materials to the synchrotron for measurement. Interpreting and understanding the data that comes out of the synchrotron is very challenging, and as a result we have started to build computer models that use quantum mechanics to simulate what happens during a synchrotron experiment. To get accurate results, these models must be systematically improved as the materials and measurements at the synchrotron get more complicated. This project specifically targets the development of computer models that can describe pump and probe spectroscopy experiments, this is where a chemical system is excited, and then we measure the properties of the excited state. Using these models to simulate realistic materials takes a long time and a lot of computational resources, the project will use GPUs to mitigate the most intensive parts of the simulations. We will use the software to analyse and explain pump and probe spectroscopy measurements of organometallic complexes.

University

University of Lincoln

Facility

DLS

Development of improved, scalable electrocatalysts for sustainable chemistry and energy storage requires understanding and eventual control of reaction mechanisms. The access to time- and voltage-resolved interface-specific insights into reaction intermediates at low concentrations makes Vibrational Sum Frequency Generation (VSFG) one of the most exciting spectro-electrochemical approaches, whose potential for fundamental research in sustainable electrochemistry has only recently started to be tapped. In spite of rapidly growing interest across the spectro-electrochemistry community, widespread use of VSFG for electrified interfaces remains hindered by both challenges in the experimental set-ups and the lack of direct, simulation-assisted interpretation for VSFG-spectra of complex and dynamic electrode-electrolyte interfaces. In parallel with ongoing experimental endeavours at CLF-ULTRA, this studentships aims to start filling the existing simulation gap for electrochemical VSFG by:

1. Development and application of new VSFG simulation capabilities to support ongoing and future research at CLF, and more generally STFC experiments on electrified interfaces.

2. Student training in simulation and software best practice in direct contact with experimental partners, contributing to future-proofing the computational electrochemistry pool of STFC (the UK).

3. Paving the way to the establishment of a common language and understanding at the SCD/CLF interface, to the benefit of CLF and SCD stakeholders.

University

University of Liverpool

Facility

SCD / CLF

Randomised numerical linear algebra (RNLA) is an exciting branch of computational mathematics, which has had a profound impact in several applications where large-scale matrix computation is required; randomised low-rank approximation of matrices is a primary example, with applications in statistics, data science, machine learning, and computational physics, etc. While the field has reached a somewhat mature stage, several open problems remain, including (i) gaps in theory and practice, (ii) randomisation for tensor problems, (iii) high-quality implementation of the best algorithms, and (iv) application of RNLA algorithms and ideas in other disciplines, for example, compressed sensing, networks, and mathematical biology. One specific example is the further study of sketching for dimension reduction, which is a key idea in most algorithms in RNLA. The project aims to develop efficient techniques for sketching a large-scale matrix or tensor. The sketches have a tensor structure allowing them to be applied faster than unstructured (e.g. Gaussian) sketches while maintaining sufficient randomness that allows algorithms to succeed with high probability. This project intends to establish theoretical justifications for such sketches and identify limitations (if any) so that we can make theoretically justified recommendations on when such sketches should be employed. We expect the new sketch to be competitive in many settings.We will keep a close eye on applications, in particular in problems involving high-dimensional image reconstruction arising in tomography. A specific application that will be investigated is the principal component analysis (PCA) for tensors, which is a key computation required when retrieving important information from a given data. Resulting sketching and PCA algorithms will be implemented in Python and GPU, and the codes will be publicly available.The project will draw from mathematical topics including NLA and matrix analysis, probability, perturbation theory, and statistics.

University

Oxford University

Facility

SCD

Muonic atoms X-ray spectroscopy is a non-destructive method for
determining the composition of a sample, which is widely used in the area of cultural heritage materials. This method involves measuring the X-rays
emitted from the interactions of the sample with a negative muon, which
is a sub-atomic particle that can be thought of as a heavier analogue of the electron. These emitted X-rays have more sensitivity to light atoms and,
due to their controlled energies, allow for depth-dependant studies of the sample. Elemental composition can be studied from micrometres to centimetres below the sample surface.
During the last two years, a collaboration between the SCD and ISIS has
been developing the software MuDirac, which is a modern, open-source,
sustainable software tool that is being used to aid with muon elemental analysis at ISIS via the computation of muonic X-ray spectra. MuDirac can
compute the energies of the X-rays emitted by most muonic atoms with
sensible precision. MuDirac, however, cannot currently calculate the intensities of these muonic X-rays, which affects its ability to perform elemental analysis efficiently. This project aims to extend the capabilities
of MuDirac by developing a dependable method of calculating the intensities of muonic X-rays. The new version of MuDirac will be made part of the Muon Spectroscopy Computational Project and made available to the muon community.

University

University of Warwick

Facility

SCD / ISIS

Multi-dimensional ptychographic imaging is becoming a popular imaging approach at synchrotron and electron microscopy facilities to obtain structural, chemical and magnetic information of complex materials with unprecedented resolution. As part of the reconstruction process, this approach combines two-dimensional ptychographic imaging with techniques such as computed tomography (CT), spectral mapping or magnetic dichroism to form higher-dimensional volumes. It is common practice to build a chain of data analysis steps, for example by feeding the results from a ptychography reconstruction software into a separate program for tomographic reconstruction or other downstream analysis tools. Even with highly optimised GPU-accelerated reconstruction codes, the end-to-end pipeline can take a substantial amount of time that cannot cope with the ever increasing data collection rates. Furthermore, the chaining of algorithms can lead to unnecessary accumulation of noise and artefacts hampering the achievable resolution and contrast in the reconstructions. In this doctoral project, we propose to build the theoretical and mathematical foundations for a new multi-dimensional ptychography reconstruction framework that combines all required processing steps into a single iterative optimisation problem and uses artificial intelligence (AI) to provide fast on-the-fly data inference that can cope with modern high-speed detectors enabling real-time feedback for users of high-throughput multi-dimensional ptychography.

University

University of Oxford

Facility

DLS / SC

We propose to deliver a total scattering tool for the improved
study of nano-sized materials with a layered structure
architecture based on numerical models rationalized by
machine-learning enhanced analysis of atomistic simulations.
Understanding the stacking of fault defects and layerstructure
units can lead to the development of materials with
improved performance in various applications, from medicine
(e.g., shape memory alloy for implants) to aerospace (e.g., highstrength
metal alloy) environmental remediation (e.g., filtration
of heavy elements), electronics (e.g., semiconductor devices),
energy production (e.g., solar cells) and storage (e.g., batteries
and supercapacitors). The Lego© model of stacking 2D-building
blocks: can alter the atom mobility to enhance the material’s
ionic/electronic conductivity, strength, ductility, and toughness.
Specific materials that we will target in this work are
turbostratically-disordered anode materials for sodium-ion
batteries (hard carbon and sodium titanates), layered double
hydroxide, used in green chemistry, and layered perovskites,
used in hydrogen production via water splitting.
To date, scattering analysis methods aimed at the study of
layered structure are limited by the challenge of capturing both
the local structure within layers and information on the stacking
sequence and faulting. Consequently, a handful of expert endusers
of Diamond and ISIS are capable of extracting significant
information from the analysis of scattering data, limiting the
range of materials this methodology can be applied to. This
project will improve the reliability and useability of total
scattering analysis software solutions/approaches. This is
achieved by designing software with the inherent capability to
bridge the length-scale gap. The code will build on the work
frame of the whole pair distribution function modelling method,
which has already proven to be able to face this challenge.
This Project is part of a larger strategic vision at Diamond and
ISIS to provide an ecosystem of software tools that enable
solutions to the problem of end-user data analysis. It is also part
of a larger international collaboration including among others
the Erlangen National High Performance Computing Centre.

University

University of Birmingham

Facility

DLS

Deep learning has has emerged as a powerful alternative to
the iterative phase retrieval approach, that can provide robust
reconstruction of Fourier-space diffraction pattern data where
iterative methods often fail to solve the phase retrieval
problem. Although emphasis to date has focussed on inversion
from Fourier-space to real-space images, the process of
recovering real-space images remains unclear due to the
inherent and currently intractable complexity of deep learning
methods. In this project we will develop Physics-Aware Super-
Resolution convolutional neural network tools to enhance the
visibility of Fourier-space diffraction patterns thus enabling
rapid and accurate reconstruction of phase information.

University

University of Southampton

Facility

DLS

Neutron spectroscopy (NS) has emerged in recent decades as a powerful technique for studying and understanding the structure and dynamics of a wide range of systems. From green energy materials, to pharmaceuticals the unique properties of NS have provided insights unavailable to other techniques, unlocking key insights and facilitating design of new targets. Molecular modelling is becoming an increasingly important counterpart to NS, with comparison of simulated and experimental results providing a direct link and allowing researchers to ‘see’ the structure of dynamics of real systems at an atomic scale. However, as NS experiments become more powerful and the systems that they can probe become more complex, access to efficient but accurate molecular modelling becomes increasingly challenging. In this project we will bridge the efficiency/accuracy divide by using machine learned interatomic potentials to drive accurate simulations of a number of important test systems. Our approach will complement developments in NS hardware and experiments, and help to unlock the potential of applying this powerful technique to some of today’s most important technological challenges.

University

University College London

Facility

SCD

A key topic in modern condensed matter physics is the search
for and utilization of a quantum spin liquid – a special state in
which there is strong quantum entanglement of spins but no
magnetic order. A crucial ingredient to realize such a state is
strong spin-orbit coupling (SOC), which leads to highly
anisotropic interactions between spins. Inelastic neutron
scattering (INS) is the ideal probe to obtain and understand such
phases. However, a break from conventional linear spin-wave
analysis is required to include the multilevel physics arising from
the SOC. We have developed an excitonic approach for this,
where we include the single ion states and the coupling
between them and calculate the neutron cross-section
numerically using Green’s functions. This approach can be
computationally expensive, making least squares optimization
of the parameters prohibitive.
We propose to develop and use machine learning-assisted
global optimization techniques to perform high-quality fits of
excitonic models to large INS datasets. The technique will
involve Bayesian neural networks, the differential evolution
algorithm, and novel techniques to integrate them. Special
objective functions beyond the traditional root mean square
method for the target problem will be investigated. The
hybridization of this with parallel computing could also be
included.

University

University of Edinburgh (lead institution)

University of Glasgow (co-lead)

Facility

ISIS

When light interacts with a material, electrons may be excited
to higher energy states. Understanding and predicting the
dynamics of these excited electrons and how they may
dissipate that energy to other particles, such as the atomic
nuclei, is an outstanding scientific problem. Experiments at
STFC Facilities, such as CLF-ARTEMIS, Diamond Spectroscopy,
ISIS-photo-muons and RFI, routinely encounter this process
when investigating a wide range of materials.
This project aims to complement these activities by
performing method development and software design to
provide a workflow to simulate electron-nuclei dynamics on
the fs to ps timescale using state-of-the-art, quantummechanical
methods without any empirical parameters. The
software to be developed will be sample-agnostic and
transferable to many different cases.
We will apply these tools to study trapped electrons in a
materials for nano-electronic devices, namely insulating oxide
films such as HfO2. Following excitation from defects states
into the conduction band, the electron may get re-trapped at
the defect site or interact with the nuclei to form polarons. As
insulating oxide films for electronics often crystalise in the
amorphous phase, which is computationally demanding to
simulate, we will also explore the applicability of Physics-
Informed Neural Networks to speed up the proposed class of
simulations.

University

University College London

Facility

SC / CLF

The Ada Lovelace Centre will fund 50% of the cost of a studentship (fees and stipend) with a partner University.

The call for proposals typically launches in June each year with a closing date of mid-late September. Successful proposals will be informed in October. The call includes more detailed terms and guidance.

Studentships can be for a maximum of 4 years.

The student is expected to spend a significant portion of their time (30-50%) with Ada Lovelace Centre teams and working with facilities during their PhD. In practice, this is typically arranged as a block of time, typically in year 2, spent with the Ada Lovelace Centre supervisors.

The aim of the studentships is to support forward looking developments and to build both the links between facilities and Ada Lovelace Centre and the wider expertise which supports our goals. This therefore requires that the PhD supervision is not simply between a University and a facility but links strongly to the scientific computing expertise in the Ada Lovelace Centre.

Any academic belonging to a UK university with the authorisation to supervise PhD students may apply as the university supervisor.