Engineering & Materials Science
Below you will find Data-Centric Engineering projects offered by supervisors within the School of Engineering & Materials Science
This is not an exhaustive list. If you have your own research idea, or if you are a prospective PDS candidate, please return to the main DCE Research page for further guidance, or contact us at dce-cdt@qmul.ac.uk
Modelling pregnancy outcomes to prevent preterm birth after exposure to air pollution during Covid19
We have an opportunity to better understand the impacts of air pollution during pregnancy to prevent preterm premature rupture of the fetal membranes (PPROM) and preterm birth. Exposure of women to pollution particles from vehicle traffic has reduced due to lifestyle changes during Covid-19 lockdown. We will compare the health of pregnancies born during the pandemic with those babies born when pollution was at its highest. This approach will give us more evidence around the impact of air pollution on the health of preterm and term babies.
Whilst many studies have found carbon particles in maternal, fetal and placental tissues, no one knows their effects on fetal membrane structure. We hypothesise particles less than 10 microns in size are the most damaging and leads to degradation of the fetal membranes and PPROM. The loss in cell function and tissue integrity damages collagen structural integrity leading to mechanical failure and early rupture of the fetal membranes.
Computational modelling and development of machine learning tools will map pollutant particle effects with fetal stage and membrane mechanics. Using AI, we will develop models that can predict which women are at greater risk of rupture and preterm birth.
Keywords: Health data, pollution, machine learning, healthcare, women's health, neonatal health.
Supervisor: Dr Tina Chowdhury
Adaptive multi-fidelity robust design optimisation driven by machine learning
High-fidelity numerical analysis is computationally expensive. The cost dramatically increases when using e.g. time-resolved unsteady simulations (LES) or including uncertainty quantification. When included in numerical optimisation loops the cost is generally prohibitive for industrial application. Heuristic methods have been developed to sequence different levels of model fidelity in the various design stages, but are difficult to extend to complex systems and novel technology.
The project will extend existing in the PI’s group with Multi-level methods for Uncertainty Quantification to a) include a much wider variety of levels of fidelity ranging from the most expensive unsteady resolution of turbulence to zero-order empirical models and b) include other disciplines such as structure, vibration and heat transfer.
The group has a long-standing experience in sensitivity methods and the particular novelty of the project is the pervasive use of cheap sensitivity information by using adjoint back-propagation through the physical models to obtain error estimates for UQ.
The project will be conducted in collaboration with industrial partner Rolls Royce who will provide guidance, references cases and industrial data.
Supervisors: Dr Jens-Dominik Mueller & Dr Andrew Buchan
Development of a new causal artificial intelligence framework for health predictors
The grand success of machine learning has led to an explosion of AI applications and an increasing expectation for autonomous systems that exhibit human-level intelligence. These expectations, however, have met with fundamental obstacles, like "robustness", that is reproducible results, “adaptivity” which is a good response in new environments, “explainability”, that is, machine learning models remain mostly black boxes unable to explain the reasons behind their predictions. It is now generally accepted that all three obstacles mentioned above require causal modeling.
In this project we explore and develop previous and newer approaches for causal AI and apply these concepts to the Health sector. The student gets access to globally obtained unique datasets from patients and experimental animals to develop new predictive models for patients. As such this project is based on large, big data from human and experimental animals, and aims to develop novel tools to predict and diagnose disease.
Keywords: Health, AI, Deep Learning, Causal AI, Do calculus, Atherosclerosis, Genomics
Supervisor: Prof Robert Krams
Trustable public transit through integrated multimodal transit networks and trip planning
This project will investigate the following questions:
1. How to identify areas where there is an increased risk of congestion or crowding in real-time?
2. How to assess areas of high risk of transport-related COVID-19 exposure?
3. How to alert travellers to stay away until there is sufficient capacity available or disperse them to other transport means?
4. How to inform service providers the available options and assist them to better manage overcrowding?
5. How to integrate active/shared/micro/mass mobility solutions into a comprehensive transport ecosystem that would let travellers navigate seamlessly across different transit modes, and help transport authorities prioritise new infrastructures?
The project will create the pathways to such technology by modelling and integrating different transit networks, assessing the risk of transport-related COVID-19 exposure, planning alternative mobility options with the assessed risk, prioritising cycle infrastructure using crowding information and informing both travellers and service providers.
Supervisor: Dr Jun Chen
Deep learning and multi-objective predictive models for drug development
The project will be divided into four stages. The initial phase will focus on materials prediction based on simulated or published experimental data. Next, the Scholar will move on to material synthesis. We expect that realisable materials will be identified and fabricated. Third, device simulations will be performed. For instance, potential radiation detectors can be studied with GEANT4-based simulations, building on predictions of electrical performance and of how the material will interact with radiation; piezoelectric energy harvesters can be modelled using finite element analysis. Finally, the newly identified materials will be tested experimentally in real devices.
Supervisor: Prof Robert Krams
Predictive modelling for 222 nm far-UVC aerosolised virus inactivation within populated indoor environments
This project will develop a predictive modelling capability that will quantify the transmission and optimise the inactivation of airborne pathogens within populated indoor environments using 222nm far-UVC light. Far-UVC has recently been proven to kill most bacteria and virus, including covid-19, but, unlike standard 254nm UVC, is shown to be safe to use around people. Maximizing far-UVC efficacy of airborne virus inactivation by a lamp's position is complex, requiring coupled physics of radiation and CFD to be considered for each room's unique conditions due their shape, size, furniture and ventilation.
Numerical simulations are extremely expensive, thus this project will build, using state-of-the-art AI techniques, a predictive model that can quickly guide the optimal placement of lamps given any arbitrary rooms conditions. This will be built within, and trained by, our unique coupled high-fidelity radiation-CFD model for assessing far-UVC inactivation of airborne diseases.
Keywords: predictive modelling, AI, covid-19, CFD, Radiation transport, optimisation
Supervisors: Dr Andrew Buchan & Dr Jun Chen
Development of a predictive thermal model for buildings through a computationally-based machine learning approach
Building heating is, currently, responsible for over 20% of the greenhouse gas emissions in the UK. Poor thermal management is an essential factor intensifying energy consumption in buildings. Although some attempts are being made to improve buildings energy efficiencies, achieving smart thermal management remains as a major challenge. This is partly because of the large varieties in building configurations. More importantly, unlike electricity, heat is a spatially distributed quantity which couples strongly with human sense of comfort, needs for ventilation and building characteristics.
To overcome the issue of complexity of building thermal management, this PhD project develops a hybrid computational and data-based model for the distribution of temperature, humidity and air pollutants in buildings. Advanced tools in computational fluid dynamics will be utilised to model the processes of heat generation, distribution and dissipation in buildings. Machine-learning will be then employed to develop a predictive thermal model of the building.
Keywords: Decarbonisation of heat; Building thermal management; Computational fluid dynamics; Machine learning; Data-driven predictive models
Supervisor: Dr Nader Karimi & Dr Jun Chen
Discovery of halide perovskite thermoelectric materials
In recent years a number of databases of computed or experimental properties of known materials have been developed. When new technologies emerge, this invites the possibility of mining data from these databases to predict the most suitable materials to realise the technology. It may even be possible to use the database of known materials to predict new ones or to propose modified compositions that improve performance. This project will focus on the development of new thermoelectric materials that can be used to convert waste heat into electricity. One particular class of these materials that has recently emerged out of research at QMUL is the halide perovskites. However, this is a vast class of materials, and experimental research could be accelerated if data mining and machine learning could be used to predict material compositions that would yield the highest performance.
Keywords: Materials discovery; perovskite; machine learning; thermoelectric
Supervisor: Dr Oliver Fenwick
Data- and physics-driven aeroacoustics modelling
Aeroacoustics is a branch of fluid dynamics which studies noise generated by aerodynamic flows. Due to the ever-increasing human impact on the environment, aeroacoustics has become an important element of the design optimisation cycle in the aerospace, automotive, and wind energy sectors. The large difference in complexity between turbulent flows and the radiated noise in the audible range of frequencies has been exploited by many aeroacoustics theories. Most recently, data-driven (data-centric) modelling has shown a significant promise to not only accelerate the data acquisition and processing but also uncover constitutive dynamics equations from the high-resolution flow simulations. This project will combine advanced data-driven methods for the extraction of useful low-order structures from complex noisy fields, high-fidelity flow simulation accelerated on Graphics Processing Units, and the aeroacoustics theory to develop a new generation of reduced-order models of jet and airframe noise.
Keywords: turbulence, computational aeroacoustics, acoustic analogy, Large Eddy Simulation, surrogate models, symbolisation, genetic algorithms
Supervisor: Prof Sergey Karabasov
Data-driven decisions for STEM students during the school-university transition
The transition from school to university is one of the most enjoyable times in an engineers life. However, the transition is synonymous with a wide range of challenges that have been demonstrated to influence long term student performance, drop out rates and overall development. While many of these challenges are supported in good quality universities, the academic dimension of the transition are lacking support in a number of key areas. In particular, engineering schools do not typically offer support to address any deficiencies in students background knowledge, skills and reasoning capabilities. Further, students from disadvantaged or diverse backgrounds are commonly at a greater disadvantage and find it harder to close any gaps. A STEM transition survey has been designed to help students identify gaps in there background and to provide resources to help them close these gaps. To date the survey has assisted around 1,000 students to identify and close gaps in there academic backgrounds.
This project aims to use and build on the large data pool collected, determine other key metric and collect additional data. The project will review and select appropriate machine learning / artificial intelligence tools with a view to understand the needs of individual students, provide more customised support and predict any risk such as failure in a specific module. The initial work will be done within engineering schools but it is anticipated to have wider applicability in the future. The project aims to produce a functional prototype which can be used to demonstrate the value of a university wide system.
Keywords: Engineering education, STEM, Data analytics, AI, Machine learning, University transition
Supervisor: Dr Andrew Spowage
Developing a novel interpretation of Optical Coherence Imaging on the basis of data-centred Artificial Intelligence
We aim to make interpretation of optical coherence tomography (OCT) images and their 3D reconstructions fully automatic by using state-of-the art neural networks with transfer learning. We have already tested the feasibility of this concept and we applied a pretrained RESNET50 model but its application was not very successful. What was learned however was that pre-processing process required improvements (catheter removal, segmentation, catheter shadow removal) and that we have to expand the data set, reduce feature space and apply other CNN models, as RESNET50 was tuned into photography.
To address these challenges we aim to add a human OCT data set (obtained from the Bart’s Heart Centre); these data will allow us to test the feasibility of this approach in the human coronary arteries in the clinical environment.
Supervisor: Prof Robert Krams
Decoding the full capacity of the information-rich multimode optical fibres through deep learning
The information-rich multimode optical fibre has found exciting applications in high-capacity communication, sensing, imaging and micro-endoscopy, and therefore are particularly attractive to combine with deep learning.
Despite recent research efforts to develop MMF based devices and systems, the information capacity of MMF has never been fully explored. For example, space-division multiplexing based optical fibre communication systems are limited to only few-mode optical fibres. For single MMF imaging probes, they have to stay rigid, as any changes to their geometry will lead to completely different transmission matrices (TM), resulting in the failed image recovery. In addition, MMF based optical sensors, are mainly intensity measurement based with limited attention on how to extract the useful information from the complex but information-rich MMF speckle patterns.
Supervisor: Dr Lei Su
Computational modelling of underground hydrogen storage by deep learning techniques
Despite the great desire for developing hydrogen economy, the storage issues continue to seriously challenge any large-scale hydrogen energy plan. Hydrogen storage technologies are usually expensive. However, underground storage of hydrogen may feature promising economics and therefore have recently attracted much attention. The hydrogen produced by the electrolysis of water is pressurised, or liquefied, and injected into the depleted oil and gas reservoirs. Such geological formations have hosted high-pressure fluids for very long periods, providing some assurance about a leakage-proof storage. Nevertheless, safety regulations still require a thorough analysis of the injection and extraction of hydrogen into/from the underground reservoirs.
The intended analysis includes simulation of flows of pressurised hydrogen into/from the porous underground reservoirs that scale from hundreds of meters to kilometres. The problem further includes extra complexities like the existence of other reservoir fluids (e.g. water), heat transfer and particularly the unique thermophysical characteristics of hydrogen. The latter disqualifies the existing reservoir models, developed for natural gas and CO2, from analysing the underground storage of hydrogen.
The building block of this highly complex problem is the flow of hydrogen through individual pores. However, realistic pore-scale analyses using numerical techniques are usually prohibitively expensive. To address this issue, our research group has recently combined the most advanced machine-learning techniques with computational fluid dynamics to build the so-called ‘Physics-informed Neural Network’ (PNN). In this approach, the physical governing equations are incorporated into a deep learning scheme resulting in a mesh-less solution of those equations. This offers a significant reduction of the computational cost and enables the pore-scale analysis of realistic subsurface flows through porous structures. The PhD includes numerical solution of Navier-Stokes equations in a 3D model of the geological formations for few cases only. A PNN is then developed that extends the solutions to a wide variety of cases.
Supervisor: Dr Nader Karimi