The CDT will enable students to develop new fundamental AI capabilities in the context of a diversity of complex systems. Rather than working in isolation, as is usual in AI, the students will learn to develop these in a collaborative manner tied to a specific application domain. In particular, the CDT will comprise three topics:
Uncertainty in complex systems (UQ). Human and AI collaborative decision making requires principled uncertainty quantification. The CDT will develop leading-edge probabilistic machine learning methods, leveraging uncertainty in a statistical manner to drive the exploration of new parameter spaces and promote scientific discovery. With a focus on the methodological and theoretical aspects, the CDT will provide impact across any field where decision-making is critical. Uncertainty quantification and modelling will underpin the following two topics of the CDT:
Decision-making with humans in the loop (DMHL). By acknowledging that the human user, in many cases, is unable to fully specify the details computer systems require, and by jointly modelling the machine learning task and the user, AI technologies will be more efficient in addressing key challenges such as experimental design from scarce data and domain shift, as well as promoting trust in AI-enabled systems.
Decision-making for ML systems (DMML). Increasingly, several automated decision-making systems are being built by a composition of individual machine learning components making decisions at the component level. For scientific systems that produce huge data volumes, so-called “big science” (e.g., SKA, CERN), AI-driven decisions are increasingly necessary to replace human decisions at multiple points within scientific analyses and facility operations. The CDT will look at automated AI approaches that can ensure such systems combination is robust, safe and accurate.
Model interpretability and explainability will be transversal to the three topics above. Decision making with AI needs to be interpretable and explainable to facilitate interrogation of decision processes such that trust can be built by the human, and it is essential for understanding and meeting ethical and legal implications.
We will start with three different types of fields of research, each allowing the cohort an opportunity to explore different kinds of new questions and identify different kinds of hypotheses with AI: Physics/Astronomy, Engineering Biology and Material Science. In Physics/Astronomy, we will work closely with collaborators at the University of Manchester (UoM) and the University of Cambridge (UoC) to bring together simulation science and ML/AI methods. The intersection of simulation science and ML/AI will drive progress in scientific experimentation and discovery and has often been driven by the “big physics” community. More widely, AI solutions must work for real-world scientific operations such as the new generation of astronomy and physics facilities, including the Square Kilometre Array and Large Hadron Collider, meaning the solutions provided must also be scalable. For Engineering Biology (EB), we envision a close collaboration with the Manchester Institute of Biotechnology in projects related broadly to human-interpretable, multi-scale predictive models that can direct the automated design and optimisation of biological systems. The ML/AI tools and methods developed through the CDT can be applied to virtually any EB application area with envisioned impacts in strategic domains of National importance: Clean Growth, Sustainable Agriculture and Bioproduction, Biomedicine, Environmental Solutions, and Novel Materials. In Material Science (MS), we will work with the Henry Royce Institute and the UK Atomic Energy Authority in ML/AI aspects of digital engineering for Materials 4.0. Among others, Royce will provide materials data repositories, and we will jointly develop physics-based modelling powered by ML/AI to accelerate materials innovation and manufacturing.
As part of the co-creation process for this CDT, we had conversations with different companies and organisations interested in the three topics that the CDT addresses, UQ, DMHL, and DMML. We will work closely with companies to ensure that the methods developed in the science domain are adapted to various business settings, ensuring real-world practical impact. An important prerequisite for this is to develop methodologies that are reproducible and accessible to all research, innovation and industry communities. To support and steer the CDT in carrying out research that translates into increasing business productivity and ultimately contributes to increased living standards and well-being, we will be working with The Productivity Institute (TPI). TPI is a UK-wide research organisation headquartered at AMBS, UoM, with UoC being one of eight partners, each partner leading a regional Productivity Forum. It includes representatives from policy, community and industry.
The UKRI AI Center for Doctoral Training (CDT) in Decision Making for Complex Systems is a joint CDT between the University of Manchester and the University of Cambridge. The CDT provides funding for four years of advanced studies towards a PhD. The first year is a taught program that will cover the fundamentals of Machine Learning. This year is followed by three years of research at either at Manchester and/or Cambridge.
Discover ‘Advancing Breast Cancer Prevention’: A research project blending interpretable machine learning (ML) with expert insights to revolutionize decision-making in breast cancer prevention strategies. Focusing on mammographic density and spatial transcriptomics, we aim to unveil deeper breast cancer causes and find effective prevention methods. This project is expected to provide promising significant advancements in personalized healthcare and cancer prevention.
This interdisciplinary PhD project pushes design boundaries in soft robotics, combining advanced probabilistic AI models with differentiable simulations to optimize materials and predict performance amid uncertainties. A crucial application of AI, driving smarter engineering of intuitively responsive and adaptive soft robotic systems through holistic integration of computational intelligence and experimental insights on next-gen materials.
This project will explore the use of a transformer-based architecture, inspired by state-of-the-art LLMs, as a scheduling method for industrial production facilities. This architecture would learn from historical data and generate new schedules in a similar way to how they generate text, implicitly learning the scheduling rules from the data. This approach will be an alternative (or complement) optimization meta-heuristics. This project will explore cutting-edge technologies from academia but with an strong link with industry.
This project will apply statistical machine learning / AI to a growing body of bacterial genomes. The aim is to facilitate the engineering of novel biosynthetic pathways for high-value drugs, such as antibiotics, in heterologous microbial expression systems. The key output will be detailed assembly rules for the design of hybrid gene clusters and for the successful domain-shuffling of multi-domain enzymes that often are responsible for the biosynthesis of natural products.
Modern radio telescopes are often built as arrays of antennas that operate interferometrically to synthesize apertures much larger than that provided by a single element. Due to improvements in digitisation, these arrays are able to be reconfigured in a flexible manner for different scientific applications, including division into heterogeneous sub-arrays, bandwidth partitioning and antenna/baseline reweighting, often for multiple sub-arrays operating in parallel. Choosing an optimal set of configuration parameters for such partitions is a combinatorial optimization problem, subject to complexities that are often specific to every individual scientific use-case. In this project the student will explore the potential of modern methods in artificial intelligence, such as deep-learning, reinforcement learning and diffusion models, to learn the heuristics of array configuration for the science operations of the Square Kilometre Array Observatory (SKAO) and its associated facilities. In doing so, this work is intended to provide insight into how machine learning more generally can be employed to solve combinatorial optimization problems on graphs.
Structural materials in a tokamak fusion reactor are designed to withstand extreme temperatures, stresses, and radiation damage during their operation. Radiation damage, from bombardment with high-energy neutrons, result in complex defect structures that evolve across a wide range of time and length scales. Quantifying this damage structure and how it relates to degradation of material strength is critical to assessing risk of failure in reactor components. While experimental characterisation under neutron irradiation is time consuming, expensive and, in many cases, impossible, atomistic simulation methods can model individual damage events, involving millions of atoms, with a high level of fidelity to the true physics. However, raw data from these simulations, described by the positions of millions of atoms, is too large to parameterize predictive models of long-term damage accumulation across reactor components. Thus, the key to unlocking reduced order models for radiation damage lies in the development of a compressed representation of the underlying defect structures. Such a microstructure fingerprint reduces model parameterization from millions of atoms to a more manageable number of descriptor features. Machine learning methods are increasingly accelerating the development of these fingerprints by treating microstructure evolution as a pattern recognition problem. In this project, we propose to generalize recent advances in microstructure image fingerprinting to graph data, which is the more natural setting to represent changes in local atomic neighbourhoods. Graph-based autoencoders will be developed to extract global and local descriptors of radiation damage microstructures in Tungsten – a critical plasma facing component which is exposed to the highest neutron doses in a reactor – using a large dataset of high-fidelity atomistic damage simulations. Furthermore, time-series data will also be used to learn the latent space dynamics of these damage descriptors. The outcome of this project will enable the parameterisation of fast and actionable reduced order models for radiation damage evolution in fusion reactor components, which will significantly accelerate their in-silico design and qualification.
Explainability of the predictions made by deep learning models is paramount in sensitive domains as well as for better decision making in general. We look at how to make deep learning models inherently explainable using a a concept learning approach.
Explore the future of engineering design with our project on ‘Generative Modeling for Engineering Design.’ This cutting-edge research combines AI and machine learning to revolutionise CAD data analysis, automating feature extraction and classification. Partnering with industry leaders Cummins, we aim to transform engineering practices, enhancing efficiency and innovation. Join us in shaping advanced AI applications in real-world engineering!
Combinatorial Decision-Making has numerous practical applications across domains varying from logistics, physics, to fundamental sciences. As the search space typically grows exponentially with the problem size, combinatorial decision-making is often tackled with handcrafted heuristics using expert knowledge. This project leverages Generative Models (GMs), including Large Language Models and diffusion models to capture those handcrafted heuristics for solving large-scale combinatorial decision-making problems.
This is a collaborative project with Honda Research Institute (HRI), who are interested in efficient and sustainable vehicle design. A typical challenge here is performing costly simulations (eg through CFD), potentially combined with resource-intense physical experiments in the laboratory, resulting in optimization problems with heterogeneous objectives and/or constraints. We will develop novel Bayesian methods for heterogeneous objectives, and validate them on real problems provided by HRI.
This project aims to develop interactive machine learning methods that help AstraZeneca scientists guide high-throughput screening in drug discovery pipelines.
Experimental Design, which is the driving force of all empirical science and much of product development, involves brute-force trial and error to reach a desired outcome. While we have seen successful applications of AI driven experimental design, they are still faced with the challenges of small data, heterogeneous inputs, and domain knowledge. This project aims to use human-in-the-loop generative models to tackle these challenges. We will test the principles in running the Design-Build-Test-Learn (DBTL) loops in Synthetic Biology through the collaboration of the supervisors.
In this project, the goal is to develop methods for tackling novel types of distributional shifts by combining development of the necessary new theory with development of methods and applying them to real cases with collaborators. The project can be tailored to focus more on theory or method development, depending on the interests of the student.
MIDAS is a generation-after-next method for theoretically complete Bayesian reasoning and decisionmaking, offering an interpretable & explainable decision at high-performance computing prices. This project aims to bring down these costs to be applied to real-time human-in-the-loop decisionmaking, working with industrial partners PolyChord Ltd and DSTL.
This project aims to pioneer interactive ML optimisation techniques to help scientists at the UK Atomic Energy Authority tackle the intricate design challenges inherent in developing sustainable fusion energy solutions and make informed decisions that can accelerate progress in nuclear fusion research.
The next decade in astrophysics presents a monumental challenge where we have to decipher vast data from new telescopes, vastly surpassing existing astronomical data set sizes. The Euclid Space Telescope, launched in July 2023, will generate a massive datasets, necessitating AI-driven decisions in scientific analyses, a trend which will continue going into SKA. This PhD project aims to carry out object classification and to leverage new AI tools to unravel galaxy formation processes and predict features. Combining astronomy and machine learning expertise, this project will develop probabilistic methods to explore new parameter spaces, pushing scientific discovery and methods. This new initiative will address challenges in astronomical datasets, offering potential for breakthroughs and the discoveries of anomalous objects that often leads to new physics. The PhD student involved will become an ML expert, whose work could impact many different areas beyond astronomy. The project, designed for collaboration within the CDT will be a test-bed for larger surveys like SKA, providing a head start for even larger future data analyses.
This project combines the expertise is Bayesian numerical methods and machine learning in the Cambridge astrophysics group, with the group’s ongoing research programme in simulation based inference. The project will develop new methods for Bayesian inference and decisionmaking via state-of-the-art neural emulators, with applications to cosmology and beyond.
The goal of this project is to look for new particles making up 85% of the matter in the universe (dark matter), utilising novel data-taking and ML techniques at the ATLAS detector at the Large Hadron Collider. Challenges include handling realistic detector environments, ensuring interpretability, and reducing energy footprints. Outcomes include insights on dark matter, innovative methods and algorithms disseminated through peer-reviewed papers and tools compliant with Open Science principles.
Pharmaceutical companies spend billions every year trying to develop safe and effective drugs. They rely on statistical methods to quantify and reduce uncertainty in their decision making processes. This project will exploit the framework of information geometry, and recently developed methodologies in Machine Learning, to address these problems. We will work with an industrial pharma lab, evaluating our novel techniques in their real data. The student will have the opportunity of an internship with the industrial supervisor, based in Switzerland.
You will conduct research at the frontier of probabilistic machine learning leveraging uncertainty in the form of credal sets. You will inspect the relationship existing between these latter and a model-free approach like conformal prediction, to ultimately discover (i) in which context one approach is to prefer to the other, and (ii) whether properties of one of the two methodologies can be used to improve on the other.
This project explores the synergy between synthetic biology and AI to overcome inherent challenges in the "Design, Build, Test, and Learn" workflow of synthetic biology. By integrating AI into manufacturing processes, the project aims to streamline, automate and optimize engineered living systems. Led by experts in both fields, the research focuses on developing automated systems for converting manufacturing tasks into lab-specific instructions. The project not only addresses current challenges in synthetic biology but also envisions an AI-assisted virtual laboratory for more efficient human-AI collaborations.
Neutrinos may hold the key to many great questions of physics and noble element detector technology has transformed the way we study neutrinos by providing very detailed images. Event reconstruction is currently one of the most limiting factors in the physics performances of these detectors and machine learning solutions have shown a lot of promise to significantly improve it. Using a novel pixel readout currently under development at Manchester, new machine learning reconstruction algorithms will be developed to demonstrate the power of 3D imaging in these detectors.
This project plans to pioneer new frontiers in A.I. focusing on solving dynamical systems in use across physics, chemistry, engineering, and cardiology. This research proposal will establish new foundations to address possibilities and paradoxes in the broad theme of AI-for-Science and thus help define the limits of feasible deep learning. The Ph.D. student trained via this program will emerge with expertise in niche areas of A.I. that are poised to become critical in the near future.
A critical question in developmental biology is how cells know when to differentiate into a new cell type. Recent advances in spatial transcriptomics provide an unprecedented increase in the spatial resolution of high-throughput gene expression measurements. In the proposed project, the student will develop machine learning methods for learning spatio-temporal dynamical models from such data to uncover the rules guiding cells to differentiate into new cell types over time.
Admission to the CDT will be made on a project-by-project basis.
Applicants must apply through the official university’s website, create an account and select the PhD Artificial Intelligence CDT program.
When asked about “Research details”, applicants can use the project details they are applying to.
When asked about “Funding sources”, applicants can choose the option “Research council”; most of the other fields in this section are not mandatory.
Regarding the Supporting Statement, it is a one or two page statement outlining your motivation to pursue postgraduate research, the area(s) of research you’re interested in, why you want to undertake postgraduate research at Manchester/Cambridge, any relevant research or work experience, the key findings of your previous research experience, and techniques and skills you’ve developed.