2021 CMP Industrial Projects

To estimate power spectral density from irregularly-sampled complex data, we are currently using a kind of generalized Lomb-Scargle periodogram (LSP). However, the theory and sampling properties of this estimator are not well understood. Appropriate theoretical background is available in Percival & Walden, Spectral analysis for univariate time series (CUP 2020), p.528ff. This project could tackle one or more of these items:

1. The LSP can be viewed as a generalized discrete Fourier transform (DFT), in other words a matrix-vector product in which row-columm dot product is the projection of the data onto a basis vector of the model. In the LSP the matrix elements do not have as many nice properties as the DFT matrix. We can speak of the Lomb-Scargle Transform (LST), of which the LSP is simply the modulus squared.
2. Check that the standard theory for DFT estimation properties (e.g. for AR(n) processes) still holds for LST. This is almost certainly the cases and this step may be trivial.
3. Given that input times are fixed, is it possible to make sense of the idea of an optimal set of output frequencies?
4. Examine tapering methods (as used for DFT) for the LST case, to determine the possible improvements to estimate accuracy.
5. Can we get estimates of autocorrelation from the LST? This would be very useful in practice.
6. (Probably hard): there is no known concept of Fast Fourier Transform (FFT) for the LST. Is anything possible in this direction? A special case would be of interest and some sort of FFT may be possible: let us allow two (and only two) time intervals between measurements in the time series data. This would be approximately satisfied by our data. (The underlying mathematics for the usual FFT involves decompositions of finite groups, so having a group theorist interested to help with this would be needed.)

This year's project aims to develop a simple hardware and software solution to capture the information from operating theatres about implanted medical devices, operations and procedures. The intention is to create a low-cost solution that could be used by the NHS to help to gather data into the new National Theatre Dataset.

The project will be run by Health Data Insight CIC and we will be able to draw on close links with expert colleagues working in the NHS and NHS Digital.

We are offering up to six intern places on this group project in 2021. Interns will work together as a multi-disciplinary team, bringing a diverse range of skills and experience to develop the project, from specification to final completion.

This project has a number of components:
1) development and evaluation of a simple and low-cost solution using existing cheap barcode scanners to capture relevant information from operating theatres.
2) exploration of ways of capturing additional information about the operation, those involved, the location of the operating theatre and relevant information from the anaesthetist - for example the time the operation started, how long it lasted etc.
3) a secure method of storing and sending encrypted information to NHS Digital so that it can be incorporated into the National Theatre Dataset where it will be linked to other patient-level information.
4) creation of resources to facilitate data collection in operating theatres - for example posters with relevant bar codes and procedures that can be scanned
5) basic analysis and visualization of the data collected

Who are we looking for?
The project will have input from clinical colleagues working in peri-operative medicine and the teams in NHS Digital responsible for collection of the National Theatre Dataset and device classification. We are looking for enthusiastic students with a diverse range of skills to be part of an innovative multi-disciplinary team: skills and experience will include topics such as hardware devices, methods for local data capture and storage; data analytics, data presentation and visualization.

What skills/experience do I need?
Enthusiasm and eagerness to work as part of a small team of like-minded individuals are the main attributes we are after. Candidates from all backgrounds are welcome, particularly if you have interest or experience in one or more of these areas:  
- use of QR or 2-d codes for information storage and transmission
- data encryption
- geolocation to room level
- Bluetooth communication
- capturing data from medical equipment (rs232 data standards etc)
- visual and science communication
- stakeholder engagement

How do I apply?
To apply, please send an up-to-date CV and a covering letter outlining what skills and attributes you would bring to the intern project and any background experience that you feel is relevant to admin@healthdatainsight.org.uk

What is the deadline for applications?
The closing date for applications is midnight on Friday 19th February 2021. After short-listing we will let you know if we require any more information and/or whether we would like you to attend an informal interview. All applicants will be notified whether their application has been successful after this process has completed.

Nowadays deep neural networks (DNNs) are widely used in speech processing applications, from hearing aids to personal assistants on mobiles phones. Conventional wisdom tells that the deeper and wider the neural network models are, the higher performance the system can achieve, and this is generally true. However, the DNNs' complex architecture can be effectively optimized in order to improve the performance as well as reducing the number of parameters. Indeed, larger models cannot be implemented in hardware with limited memory and computational power. Therefore, optimizing the structure of DNNs is an active research direction.

Various methods have been proposed to optimize DNN architecture and for model size reduction, such as pruning redundant models or exploiting the sparsity of rectifier activation function (ReLU) to reduce the computational load of convolution [1]. The ReLU activation function enables a network to easily obtain sparse representations [2].

Achieving sparsity in neural networks is one of the necessary conditions that the models can be reduced in size while maintaining the same level of performance by eliminating the zero weights after training the DNNs [1]. Sparsity in DNNs can be achieved by various approaches, for instance by using sparse evolutionary training (SET) algorithm [3].

In this project, we will investigate the way to achieve sparsity by studying activation functions for DNNs. More specifically, we study the use of splines in the activation functions of a deep neural networks. Spline-related parametric models were studied to optimize the shape of neural activation units [4, 5, 6]. The use of parametric splines makes it possible to establish a direct connection between training DNNs and activation functions and the resulting sparsity. In [5], the author showed that the optimal network configuration can be achieved with activation functions that are nonuniform linear splines with adaptive knots. The study’s significance is that the action of each neuron is encoded by a spline whose parameters are optimized during the training procedure. The proposal resulted in a computational structure that is compatible with deep-ReLU, parametric ReLU [7], and MaxOut structure.

In our work, we will focus on using sparsity as one of the constraints for training DNNs with parametric activation function, in particular using the spline model proposed in [6]. Achieving sparsity could result in improved performance and sparse weights which is useful for performance improvement and model size reduction.

The student will work in a team. Besides the main supervisor (Dr. Cong-Thanh Do), there will be a co-supervisor (Dr. Rama Doddipatla) to which the student can talk to about the project.

The working hours of the lab are 9am-5pm. Given the current situation regarding the coronavirus, working remotely is acceptable. Access to the office could be considered if necessary and according to the situation.

GSK is a FTSE100, science-led, global healthcare business currently ranked as the leading pharmaceutical company in the UK. Our research is focused on immunology efforts including small molecules, biologics and vaccines. Never has it been more important for us to reduce the time it takes from identification of a potential therapeutic to a marketed medicine, that is the main focus of this project.

At GSK we have created a world-leading data and computational environment to enable large scale scientific experiments that exploit GSK's unique access to data. Our focus is on bringing data, analytics, and science together into solutions for our scientists to develop medicines for patients. A key enabler of this effort is the ability to extract knowledge from imaging data. A specific challenge for the early phase of drug discovery programmes is to assign potential drug molecules into those with a desired effect of the drug target in mind and those with an undesirable effect, e.g. toxicity.

One way to achieve this goal is by developing advanced image analytics algorithms, where image sets of cells in the presence of molecules with known undesirable mechanisms are used to define image signatures, the so-called "ground truth". Thereafter the algorithm is applied to unknown compounds to allow us to focus on compounds that are free from potential liabilities, thereby improving drug failure rates and overall speed up the often lengthy and costly drug discovery process.

In recent years GSK has invested in the creation of a Data Lake, an infrastructure where all the data generated in the company is stored and made available. This has many advantages, as the data is not scattered across data silos, and it is generated using a standardized and controlled process.

One component of this Data Lake infrastructure is the pipeline for the sequencing of genomics and transcriptomics data (RNA-Seq and other technologies). We have built a process to generate and curate this data using standard tools and parameters, generating a high-quality dataset from experiments executed from different departments in the company.

The scope of the research project will be to develop methods for meta analysis of the genomics and transcriptomics data in this dataset, comparing experiments generated by different units and collaborators.

The student will explore methods for meta-analysis of sequencing data from different experiments. The desired outcome of the project will be a computational notebook or a script documenting recommendations for meta-analysis methods, taking in consideration existing literature, and using example data from our dataset.

This project is relatively open-ended and the student will have space to explore different solutions, as well as working with a curated dataset. Knowledge of NGS is not required although some preliminary understanding may be useful. Preferred programming languages would be R and Python.

- Leek et al, Nat Rev Gen 2010. Tackling the Widespread and Critical Impact of Batch Effects in High-Throughput Data
- Evans et al, Brief Bioinf 2018. Selecting Between-Sample RNA-Seq Normalization Methods From the Perspective of Their Assumptions
- RPKM, FPM and TPM clearly explained https://www.rna-seqblog.com/rpkm-fpkm-and-tpm-clearly-explained/

In daily life we use consumer products: household cleaning products, anti-perspirants, hairsprays, etc., that produce unintentional exposure to chemicals. Unilever assures consumer safety by assessing the toxicity risk of every ingredient in every product. 

Current risk assessments use historical rodent studies [1]. Next Generation Risk Assessments (NGRAs) use New-Approach Methodologies (NAMs) that leverage advances in computing, genetics, and statistics – novel in silico and in vitro approaches that assess risk without testing on animals.

Over the past few decades, Unilever’s Safety and Environmental Assessment Centre (SEAC) has worked closely with Industry, Academia, and Regulatory Bodies to develop a wide range of non-animal approaches using mathematical modelling [2], [3], cell culture-based experiments [4], [5], and omics for systemic and local toxicity [6].

The aim of this short project is to help advance current inhalation risk assessment methods.

The overall goal of any Next Generation Risk Assessment is to get from an in vitro Point of Departure (PoD) to a relevant in vivo exposure level – getting from exposures (or concentrations) of chemicals with no toxicological response in cell models to exposures with no toxicological response in consumers and vice versa.

Obtaining an in vitro PoD from an inhaled in vivo exposure requires an Inhalation Dosimetry Model. Inhalation Dosimetry Models calculate the fraction of the total number of inhaled particles deposited in the lung – a mass per volume metric. The in vitro PoD is obtained by distributing the number of deposited particles over the surface of the lung [7].

Current industry standards use the free but closed-source Multiple-Path Particle Dosimetry (MPPD) Model [8]–[10] to calculate the deposition fraction. The MPPD Model is user friendly and well-tested. But its dynamics are based on laminar fluid flow in a pipe – the lungs are modelled as a series of branching connected tubes. This means a lot of complex phenomena: wall impacts, convection, turbulent mixing, etc. are absent from the model. 

The problem statement follows – Investigate and quantify in-vitro PoD uncertainty due to in-vivo modelling simplifications / approximations.

Advances in computational power have meant that advanced fluid dynamics simulations can be performed on most computers. Consequently, the modelling approaches of previous generations need to be revised. A significant amount of effort has been invested in advanced aerosol modelling for pharmaceutical companies, with the results published in the public domain and the software released as open-source [11].

Given the current SoA we see 3 potential directions of investigation; the student is free to tackle the problem as he/she sees fit.

1) Data-driven Approach: review the results from the current State-of-Art (SoA) in aerosol deposition simulations and build a data model to estimate the uncertainty in deposition fraction estimates. Currently available data may differ in particle size, composition (dust vs droplet), breathing patterns (smoking versus unintentional inhalation), etc. Given the available data, the student is free to use a variety of statistical techniques to achieve the required goal.

2) Computational Approach: AeroSolved is an open-source computational fluid dynamics model based on the open-source OpenFoam code base [12]. Its features include simulations of mass, momentum (Navier-Stokes equations), and energy conservation equations; Multispecies formulation for gas (vapor), liquid (droplet) phases and solid particles; and advanced aerosol physics models. Benchmarking the MPPD Model against the Eulerian AeroSolved aerosol deposition Model will provide a numerical estimate of the deposition fraction uncertainty.

3) Analytical Approach: The Equations governing the laminar-flow transport of aerosols in the MPPD Model are known and can be compared to the more complex Navier-Stokes transport model. Using the characteristic scales of the problem, a scale analysis of the (non-linear) term difference and their corresponding transport equations will yield an upper bound on the uncertainty of the deposition fraction. More generally, derive a scaling Equation for the deposition fraction for laminar and turbulent fluid flow.

The aim is to determine a numerical estimate of the uncertainty in Deposition Fraction calculations in Laminar Flow Deposition Models, if possible, by benchmarking against State-of-Art Deposition Models.

Silvaco develops Electronic Design Automation (EDA) and Technology CAD (TCAD) software. One of the modules is the tool for simulating stresses inside the electronic devices (caused by bending, heating, internal stresses etc. during device manufacture).

Your task would be:
- research the standard stress problems for which analytical solution is available.
- set up the problem within Silvaco Stress Simulator -- compare the known analytical solution with the results obtained with simulator.
- optionally, you may also set up a simple problem which can be solved numerically using alternative tools (e.g. Matlab) and compare these results with Silvaco. Depending on results of the project it may result in academic paper or conference publication.

We're looking for talented individuals to join our Research team as Summer Research Analysts to help us achieve our mission. You will contribute to the development of innovative research. We don't expect you to have detailed knowledge of the asset management space, though experience in Finance or Economics is useful. What we are looking for is an intellectual curiosity that will drive you to learn as much as you can whilst you're here, together with a background in quantitative analysis and strong programming skills.

To give you a little background, BlueCove is a scientific asset management firm founded in 2018. Here, we believe that scientific active investing, as an alternative and complement to both passive and traditional active investing, is set to be the next defining development for the fixed income industry. As one of our Research Analysts, you will join us for 12 weeks from 14th June to the 10th September 2021 (but we can be flexible on dates and the length of the programme). As well as learning about scientific fixed-income products and the asset management industry overall, you will spend your time working on an important research project.

The BlueCove Research team have a number of interesting projects to undertake. The research projects our Summer Research Analysts undertake are likely be in the following areas:
- Natural language processing (NLP) techniques applied to company-level datasets. Goal/outcome: Improving sentiment signals for use in our fixed income systematic strategies
- Analysis of Environmental, Social, Governmental (ESG) datasets to determine quality, similarity to other data sources, & usefulness for investment decisions. Goal/outcome: A better understanding of the pros/cons of various ESG data sources
- Construction of proprietary analytic measures for derivative instruments (such as options on bond futures). Goal/outcome: A set of thoroughly tested return & greek exposure calculations which can be used to form the basis of a systematic investment strategy.

We are flexible on whether you work remotely, in the office, or a hybrid of both. We compensate our interns at a market-level salary for their valuable work and you will be eligible for benefits including private medical health insurance & a virtual GP.

As one of our Summer Research Analysts, you will also benefit from:
- Weekly firmwide meetings, in addition to team meetings, so you'll learn more about the business overall
- Regular firmwide colloquia and access to other learning & development opportunities
- Agile & flexible working
- Our Employee Wellbeing Programme
- Modern technology
- The chance to broaden your network through regular firm and team work and social events
- Our modern office which is ranked "Excellent" by BREEAM
- No prescriptive dress code
- A generous holiday allowance

https://papers.nips.cc/paper/1866-using-the-nystrom-method-to-speed-up-k...
https://people.eecs.berkeley.edu/~brecht/paper/07.rah.rec.nips.pdf
https://stanford.edu/~jduchi/projects/SinhaDu16.pdf

A popular Chinese card game requires 4 players; each with 13 of the 52 cards. The goal of the game is to arrange the 13 cards in 3 sets of 3, 5 and 5. Each set is then compared with the corresponding sets belonging to the other players, and the best set in each group wins.

In this project, we want the student to investigate one or more of the following questions:
1. Performance vs computational complexity of a hard decision logic based AI
2. Performance vs computational complexity of a deep reinforcement learning based AI
3. How accurately can we predict our chances of winning based on the information that is already revealed?

MM has grown steadily over the last 14 years into a multi-million-pound business. As with any rapidly growing business, it is essential that with increased sales, consistency of quality and service level must be maintained and increasing scale requires progressive thinking and novel approaches to succeed. Cut flowers, whilst superficially a non-essential product, are an emotionally driven purchase and therefore the highest standards must be maintained. Historically the fresh produce industry, including the flower industry, has been slow and inefficient in using the vast amount of data that is generated to inform decisions within their businesses. MM recognised this need several years ago, and as it's growth has continued, so the need to generate and use data to inform sound decision making has intensified. The COVID-19 pandemic has intensified this need even further. During 2020 the demand for online purchasing has skyrocketed as consumers change their shopping behaviours to avoid contracting the virus. This presented numerous challenges and opportunities for MM to shape and adapt the next stage of growth. Data has been and will be at the heart of successfully adapting to the constant changes that all of us are subject to currently. Insight from data collection, enables continuous improvement, basing both strategic and tactical approaches on quantifiable data and robust analysis, which should ultimately allow MM to continue its growth trajectory successfully.

An example of how MM ensures product quality is through the operational department meeting required standards for bouquet production. A dedicated quality team undertake daily inspections of the flowers from the point of receipt and throughout the manufacturing process to support the operational delivery. MM also utilises it own dedicated R&D business, APEX Horticulture, to complement the operational efforts and help provide solutions to maintain or enhance flower quality.

Whilst APEX and the intake quality function have long term, established data sets, dedicated data collection during bouquet manufacturing is a relatively new venture, requiring insight and improvement from a talented student with fresh ideas and an aptitude for data analysis. As such, there is the possibility to develop processes to allow for future data to be incorporated and analysed more efficiently across the Technical function (and indeed the wider business), allowing for quicker and more accurate decision making. Whilst this placement will focus on production, the scope of the project is wide, with opportunities for the successful student to improve processes for data collection and analysis across both quality control and quality assurance.

During this project, the prospective student will have access to extensive data, including quality assessments on receipt of the flowers, operational quality assurance, and retailer consumer metrics, for example. These datasets present an opportunity to undertake more detailed analysis of long-term trends, and how various factors influence the end consumer quality and performance. Previous CMP placements within the business have highlighted several trends and added real value to the business and changed our working practices; this is therefore an opportunity to make a real difference and see your work applied in industry. The successful student will spend their placement within the technical department, which focuses on quality and customer relations, but will experience all aspects of a fast-paced, fresh produce business. This placement will also include liaising with different departments, project management, communication skills, and working towards the needs of the business.

Skills Required: Strong computer skills, Experience with statistics and modelling, Clear communicator, Self-motivated, Demonstrates initiative, Project management, Problem solving, Industry focussed.

What you will do:
- Develop an understanding of quantum algorithms and industrial applications of quantum computers
- Research, devise and develop algorithms and software to enhance Riverlane's capabilities
- Contribute to one or more projects that are core to Riverlane's scientific goals
- Discuss ideas with colleagues and communicate research in the form of presentations and reports

Requirements:
- A current undergraduate (at least third year), master's student or PhD student in a highly numerate subject, such as a science, mathematics, computer science, engineering, or a related technical field
- Experience with at least one programming language
- Excellent critical thinking and problem-solving ability
- Strong communication skills, both written and verbal
- Ability to take initiative and to work well as part of a team

Please visit our website (https://www.riverlane.com/vacancy/quantum-computing-summer-internship-sc...) for more information and to apply.

Our full-time summer internships are designed to enable current students in a quantitative field to translate their skills and expertise into an industrial setting. You will join us at our office in Cambridge, UK, for 10 to 12 weeks, where you will have the opportunity to work alongside our team of software developers, mathematicians, quantum information theorists, computational chemists and physicists - all experts in their fields. Every intern will have a dedicated supervisor and will work on a project designed to make the best use of their background and skills whilst developing their knowledge of quantum computing. We will support all interns to try and produce a concrete output by the end of the internship e.g. a paper, product, or software tool.

We will consider remote internships depending on the Covid-19 situation.

Faraday Predictive is a Cambridge-based business, founded in 2017, with a world-leading technology for remote monitoring and diagnostics of industrial machinery that not only improves customers' business performance, it also contributes to reducing climate change impacts.

This technology is based on a whole suite of mathematical techniques, many of which have been developed in close collaboration with the Maths Department, directly involving students undertaking CMP summer projects. The students' contribution has had a significant impact on the success of the business. We continue to develop further improvements in the technology and its capability - and hope that this year's student(s) will see their projects have a similarly significant impact and benefit.

This project is all about identification of patterns in machine behaviour, to allow for identification of deviations from normal. The patterns of interest are in the form of spectral shapes, which our system creates for each machine every time it takes a reading - which can be several times per minute. So we have many many spectra as a basis from which to work - sometimes hundreds from one machine, which may or may not vary through time, and different spectra for each different machine that is monitored (and there are many machines).

Each machine has a "natural" spectral shape when it is in good condition. If a fault starts to develop in the machine, the spectral shape changes in some way, and we use this change through time to trigger warnings, and to diagnose the nature of the fault, allowing us to provide specific advice on recommended corrective action, and how soon it should be executed.

But if the first time we ever take a reading on a machine, there is already a fault present, we want to be able to identify this as a pre-existing fault, and not simply accept it as normal for this machine. At present we do this by comparing the shape for this particular machine against a spectrum for a "typical" machine - which is actually a simple averaged-out spectrum from a wide range of machines of different types.

Because the natural spectrum shape of each different type of machine is different, this "typical" spectrum is not a very good basis against which to decide whether the new machine that we are seeing for the first ever time has any faults or not.

Instead, we would like to create more specific "normal" spectra, for each type of machine, or group of types of machine, allowing us to select the one most appropriate to the machine in question.

So the tasks that we envisage being involved in this project are:
1. Create a simple algorithm for deciding closeness of fit of one spectrum shape to another.
2. Run this algorithm on the spectra in our databases to identify patterns of similar shapes - which we might call groups of machines
3. Cross refer these groups against known physical characteristics (eg machine type, duty, speed, size etc) to identify parameters that define what group a (new) machine would be expected to fall into
4. Create "normal" spectrum for each group (maybe just a simple average of them all; or we may want to take into account any known defects in some of the samples, so they are removed from the normalisation process)
5. (if time permits) - come up with a measure of "abnormality" - how far does any one spectrum deviate from the "normal" spectrum for the group in which it has been placed.

A successful outcome of this project will be that we end up with identification of patterns and sensible comparisons allowing us to provide more precise diagnostics, more precise indication of "normal" vs "abnormal", particularly for the first time we ever take a reading on a new machine, but also for application during subsequent changes through time.

Typical high field superconducting magnets can have a magnetic stored energy of around 10 MJ or more, which is 5 times the kinetic energy of a 4 tonne HGV travelling at 70mph. In the superconducting circuit, as there is no resistance, the current flows without energy loss. However, an extremely small disturbance, say 10 μJ, can lead to the superconducting magnet windings becoming resistive locally. This leads to a chain reaction, with the whole magnet becoming resistive and the stored energy dissipating as heat in the magnet windings over a few seconds' time scale. This is known as a magnet quench.

There are various schemes to protect a superconducting magnet against the effects of the rapid stored energy dissipation in a quench. The main goal is to prevent too much of the energy from being dumped locally which can produce a hotspot within the magnet windings. Typically, a protection circuit will consist of a potential divider network with resistors and diodes to manage currents and voltages within individual parts of the magnet. It will often also include secondary heaters to spread the quench across other parts of the magnet faster than the passive quench propagation would proceed.

Oxford Instruments has recently proposed a novel quench protection scheme involving varistors which is the subject of a patent application (not yet published). A varistor is an electrical device which exhibits a non-linear voltage vs current relationship. Specifically, at low voltage a varistor has a relatively high electrical resistance which decreases with increase voltage. Modelling of the quench behaviour and some experimental work has shown that the use of such components could be useful in a particular configuration to improve the quench protection for high field magnets.

Analytical equations can be derived based on the underlying physics using reasonable approximations. However, even for the simplest case of a homogenous coil divided into two sections and protected using conventional linear resistors, the result describing the propagation of a quench through the magnet coils is a second-order non-linear ODE for which only approximate solutions can be found with some further assumptions. It is not clear how to find solutions of the ODE representing the generalised case with variable resistance in the protection circuits.

Although numerical simulations for this system could be developed, analytic descriptions of how varistors would respond in a quench protection circuit will be invaluable in providing insight into the behaviour of the system over a wide range of parameters. This will also enable additional functionality in existing in-house software without requiring a great deal of computational resource.

The primary goal of the project is to find approximate analytical solutions to the equation representing the magnet quench propagation in the simplest case of the protection circuit subdividing the magnet into two sections, but generalised to allow for the use of varistors. The varistor behaviour may be represented as a simplified equation, but it may be possible to extend the treatment to allow for a more accurate representation. The use of numerical solutions to guide and test in the search for approximate analytical solutions would be appropriate, and such solutions would still be useful should analytical solutions prove to be too elusive.

This project would advance Oxford Instruments' understanding and modelling of varistors for use in protecting superconducting magnets. It is intended that it would thus support development of their practical use in the manner described in our patent application by helping in the selection of materials parameters required for a real magnet. The implementation at Oxford Instruments is likely to be in two ways: via an analytical tool to make initial estimates and by using the equations in our in-house quench modelling code.

If the project were particularly successful, an extension goal would be for the student to start working on these tools.

An academic outcome would be at least one published paper, possibly in a mathematical physics journal, but more likely a magnet/physics one.

Thermal oxidation of silicon is a way to produce a thin layer of oxide on the surface of a wafer in the fabrication of microelectronic structures and devices. The technique forces oxygen to diffuse into the silicon wafer at high temperature and react with it to form a layer of silicon dioxide: Si + O2 -> SiO2. The oxide layers are used for the formation of gate dielectrics and device isolation regions. With decreasing device dimensions, precise control of oxide thickness becomes increasingly important. In 1965 Bruce Deal and Andrew Grove proposed an analytical model that satisfactorily describes the growth of an oxide layer on the plane surface of a silicon wafer [1]. Despite the successes of the model, it does not explain the retarded oxidation rate of non-planar, curved silicon surfaces. The real cause for the observed retardation behaviour is believed to be the effect of viscous stress on the oxidation rate [2-3].

In this project, we aim to explore the existing mathematical models of the stress-dependent oxidation and propose a numerical scheme to obtain the solution. The approach that we will use is a combination of analytical and numerical analyses of a system of non-linear ordinary differential equations. The student is expected to implement the numerical algorithms in C++ language, although no previous experience in C++ coding is required. Silvaco's own software products may also be used as a tool in this project if required.

The aim of the study is to use "deep learning" (a form of artificial intelligence) to determine whether patients that fall within a particular "multimorbidity" subgroup are in greater need of healthcare services in future (e.g., more frequent doctor visits, prescriptions, hospitalisations, etc). The MSc student will contribute to the creation of a "proof of concept" for the above study question that will be used to help inform future decision making and planning of next steps on the project.

The student would have an opportunity to:
- Learn about and apply deep learning and artificial intelligence
- Apply these techniques to a real-world database (e.g., MIMIC or CPRD)
- Interpret the outputs of the analysis in a meaningful way to support scientific decision-making at AstraZeneca

The student would split their time between the above project and working on other "live" projects running within the Data Science and AI team, providing students an opportunity to work on a wide variety of tasks that a data scientist typically face during a normal working day.

AstraZeneca is investing in this exciting and vital area of drug development to generate unique multi-omics data sets to accelerate the development of novel therapies. Several projects are currently in progress to integrate microbiome with heterogeneous data sets (imaging, multi-omics, clinical, in-vivo disease models, etc.) using quantitative approaches. Collectively, such a system could lead to new targets and unique signatures correlated with human diseases. The collaborative study of altered microbial taxa/species and corresponding clinical phenotype by compiling a large and diverse data set will be an essential step toward understanding microbes' role in disease comorbidities. To achieve this goal, we collaborate with Microbial Sciences across a portfolio of projects that span multiple disease modalities.

The student will develop multi-scale models capable of integrating multi-omics data with clinical and imaging data using modern machine intelligence methods. The incoming candidate will be part of the Special Projects and Research Team. The team is currently working on a portfolio of projects with a common goal of accelerating drug or target discovery using machine intelligence methods. We aim to cross-train the incoming student in drug discovery, precision medicine, multi-scale biology, and data science. We expect the student to leverage high-performance computing and biomedical informatics facilities in AZ to develop data-driven methods to analyze large multi-scale, multi-omics data sets. The student will be part of collaborative efforts across microbial science, artificial intelligence, and drug development. This unique collaborative nature of the project will improve hands-on skills in clinical data, biomedical data analytics, and data science. The incoming student will contribute to the design, development, and deployment of predictive models that help organize, analyze and interpret. The student can also gain experience by working closely with the Microbial Sciences clinical development team.

Multiple graph modeling initiatives are running in parallel and this project will leverage their infrastructure, graph modeling of external clinical and biomedical data as well as expertise. In collaboration with this wider community, the TrialGraph project will seek to leverage these resources while developing novel graph representations of historical AZ trials, methodologies to analyze these graph representations that provide meaningful insight and experiment with other machine learning methodologies that could yield both novel discoveries and operational efficiencies.

Expected Outcomes:
- Prototype graph data mode lapplied to multiple clinical trials
- Graph analytics aimed at providing insight into clinical trial operations and outcome
- Improve clinical trial enrollment lifecycle

We would like to formulate an open-ended project made up of two parts:
* A literature review
* A practical example.

The literature review should provide a feeling for the state of the art in this area. In particular, it should point us towards what are the most promising current applications of QML to Pharma.

The practical example should be chosen based on the results of the literature review. It will be dimensioned according to the available time and QC resources available. Data sources may include publicly available clinical datasets, text, images and/or genomic sequences depending of the selected application.

There are a number of QC providers in the market place at the moment but for this purpose as well as for the literature review, it would be very interesting if we could set up a 3-way collaboration between the Cambridge Math Department, The QC group in Cambridge and AstraZeneca. This relationship could then be continue beyond this student project.

Ahren Innovation Capital is an investment fund with a remit to invest in and help build transformational companies at the intersection of deep technology and deep science that will have a positive impact on the world.

Ahren's broad fields of investment activity include: Brain & Artificial Intelligence; Genetics & Platform Technologies; Space & Robotics; And Planet & Efficient Energy. Whatever the domain, Ahren believes in taking asymmetric, considered risk that will deliver superior rewards -- capturing a generational opportunity to provide smart capital to deep technology.

Unlike in public markets where metrics of company health are established and quantified, the privately owned start-up companies that Ahren invests in have historically required a manual and qualitative approach to assessing company health and potential to create outsized returns.

This is largely a data problem. In public markets, where companies are legally required to publish their financial and operational results, the volume of data available for automated analysis is plentiful, consistent, and constrained to relatively few sources of truth in a structured format. On the other hand, private company data is rarely publicly disclosed, and the small sample of data that is shared is typically unstructured or semi structured and spread unevenly across many resources and data types (numerical and text). In some of the most exciting cases, where companies operate under the radar in "stealth mode", there is very little information at all.

As Ahren, we seek advantage in overcoming the historical constraints to quantitative early stage investing by designing novel, complementary systems to enhance our deep domain expertise in the areas that matter most.

Key drivers of Ahren's success is its ability to rapidly identify and assess world leading commercial technologies and gaps within industries that are ripe to have their biggest challenges addressed by innovation. Therefore, Ahren is starting with the automation of those tasks.

Project Goals:
1.To produce an automated Industry Model to address the following: "Help me to identify which sectors within an industry could produce attractive investment opportunities for Ahren"
2.To produce an automated Technology Model to address the following: "Help me to find the top five academic groups, globally, for any given research area."

For each model, Ahren has set out a non-exhaustive list of high-level questions to be assessed using the cross-domain expertise of Cambridge's Statistical Laboratory and Ahren Innovation Capital:

Industry
- How healthy is the industry?
- Are there favorable industry dynamics [automated Porter Analysis]?
- Can the industry be disrupted?
- Which sub sectors are receiving greatest capital allocation and seeing the most exits?
- How big could this become?

Technology
- Which are the top academic groups in this domain?
- Did this technology originate in a top academic group?

This project will require originality and creativity, bringing to bear the potential of mathematics, statistics, and machine learning to collate and derive insight from semi-structured and unstructured data. It is essential that a quality data set is built. The data set should be kept up to date and relevant using application programming interfaces and web-scraping techniques.

There are many sources of data and a good project will use a range of sources. A non-exhaustive list of possible sources is below:
- Relevant industry resources (CB Insights, Pitchbook, Beauhurst)
- Employee type / growth (LinkedIn)
- Patent applications (Escape net) or licenses
- Github / (specialized blog) activity
- Founder / individual blog posts on topic within sector
- Company / Founder social media posts and following
- Publications / conference posters: Quantity (PubMed or equivalent) and Quality (journal impact rating, paper awards)
- News articles (e.g., using news scraping tool such as Factiva)

This interdisciplinary project would ideally be completed by one Mathematician and one Computer Scientist.

The project would be of an interest to a student considering pursuing a career in investment management. Covariance matrix is a central tool in the estimation of the risk and so its estimation and filtering is very important. We would like to review the different techniques and their robustness in regards to the dimensionality, sample size. The goal is to separate the noise from real signal and filter/interpolate/update it in time.

Examples of techniques to review could be Random Matrix theory, different techniques to average covariance matrices. Naive arithmetic mean of Symmetric Positive Definite matrix conduct to swelling effect. Geometric mean on SPD matrix is one technique to cope with it.

Considering the vast array of techniques, the student will have to be critical about the benefits of one technique over the other. We will first test the techniques on synthetic generated and plotted data in Python and then test them on real world one.

Recent development in Machine Learning showed that Deep Learning methods could be applied efficiently to calibrate an option pricing model. During that internship, we will focus on two approaches, namely the historical calibration [1] where model parameters are estimated from historical market data and the volatility surface calibration [2] where parameters are obtained by inverting the market implied volatility surface. Those two problems will involve latest development in machine learning area such as the use of the signature function [3] or the Swish activation function [4].

To apply: https://jobs.brassring.com/1033/ASP/TG/cim_jobdetail.asp?partnerid=25078...