PhD: Real-time data assimilation to correct model bias in multiphysics flows

PhD: Real-time data assimilation to correct model bias in multiphysics flows

Research cut-off: 6 November 2026
Assessment: Promising with material uncertainties

Quick Decision

AE0095 is a 3.5-year Imperial Aeronautics studentship with Dr Andrea Nóvoa on bias-aware real-time data assimilation for multiphysics flows. The project starts from a practical mismatch: physics models extrapolate but have uncertain parameters and missing physics, while measurements are sparse and noisy. The advertised work is to estimate and correct model error as dynamics evolve, learn from new measurements without violating physical constraints, prototype on low-order systems and scale toward higher-dimensional multiphysics problems. [S01]

The official page states full tuition coverage for Home, EU and international students, £23,805 annual tax-free stipend and a 4 January 2027, 23:59 UK deadline. The start window is 1 August 2027 to 1 July 2028. Applicants need a first-class MEng/MSci or equivalent in a mathematical or engineering field such as Aeronautics, Mechanical Engineering, Mathematics, Statistics or Physics; computational physics/mathematics, data assimilation and dynamical-systems experience are advantages. [S01]

The project has strong supervisor-team evidence: recent papers address real-time thermoacoustic assimilation, model bias, sparse-data prediction and data-assimilated learning. The central uncertainty is transfer from low-order and selected demonstrator systems to high-dimensional multiphysics. No specific physical system, observation modality, benchmark, performance target, computing allocation or work-package timetable is public. Applicants should clarify these points and distinguish the project’s advertised objectives from the established group’s prior results. [S15–S23]

Research Problem and Likely Doctoral Work

Why bias-aware assimilation is needed

Model-based forecasts can represent physical constraints and extrapolate beyond observed conditions, but their equations may omit processes and their parameters may be uncertain. Measurements inform the true system but may arrive sparsely and with noise. When conditions change quickly, bias can grow and forecasts degrade. AE0095 proposes methods that estimate and correct model error while the system evolves, assimilate measurements while learning, and retain physical constraints. The stated progression is to prototype on low-order dynamical systems and scale toward higher-dimensional multiphysics problems. Examples in the advertisement include gusts affecting a wing and sudden combustor-load changes; these illustrate changing conditions, not a promise of a specific application or experimental setup. [S01]

Research questions

The central questions are how to represent model bias separately from state and parameter uncertainty, how to learn from sparse/noisy observations without losing physical consistency, and what evidence is needed to show that the method remains useful as dimensionality and physical coupling increase. A focused project conversation should clarify which model family and observation types are in scope, whether the target is estimation, forecasting or both, and how the low-order prototypes connect to later multiphysics demonstrations.

Recent Research and Field Direction

• Nóvoa and Magri (2022), “Real-time thermoacoustic data assimilation.” Develops bias-aware Bayesian ensemble assimilation for low-order thermoacoustic models and real-time measurements. It is direct group precedent but is not the high-dimensional AE0095 deliverable. [S19]
• Nóvoa et al. (2024), “A real-time digital twin of azimuthal thermoacoustic instabilities.” Combines a low-order model and sparse measurements to infer state, parameters and model/measurement biases. It illustrates one specialized application and cannot establish AE0095’s multiphysics scaling. [S20]
• Ozan and Magri (2024), “Data-driven computation of adjoint sensitivities…” Studies data-driven sensitivity estimation in thermoacoustics, adjacent to learning and correcting imperfect models. It is a separate approach. [S21]
• Ozan et al. (2025), “Data-assimilated model-informed reinforcement learning.” Integrates approximate models, sequential assimilation and control-aware learning for a partially observed chaotic system. It shows a nearby research direction, not the AE0095 objective. [S22]
• Özalp et al. (2025), “Real-time forecasting of chaotic dynamics from sparse data and autoencoders.” Combines latent representation, a forecast model and sequential assimilation. It is recent team research, still a preprint and demonstrated on mathematical chaotic systems. [S23]

More broadly, data assimilation is moving toward hybrid physics/ML methods that update model error and uncertainty online, and toward reduced representations for high-dimensional systems. Important open challenges include observability, calibration, stability under distribution shift, computational latency and credible validation. The project’s low-order-to-high-dimensional plan addresses a meaningful gap, but the scope and evaluation strategy must be confirmed with the team.

School, Programme and Research Environment

The degree is an Imperial Aeronautics research PhD, advertised for 3.5 years. Department groups include physics-aware machine learning and data assimilation, flow control, turbulence and computational fluid dynamics. Imperial reports Aeronautics had the UK’s highest proportion of world-leading researchers in REF 2021; this is a department-reported REF claim, not a current discipline league-table ranking. [S02–S04, S10]

Imperial’s Data-Driven Engineering pages describe research in real-time learning, bias-aware assimilation and flow modelling. The named supervisor’s group publication record demonstrates activity in this area. Department and College pages list Linux/GPU systems, central HPC and research-software support. Access requires registration; the project advertisement mentions HPC but does not specify a resource allocation, GPU configuration or queue priority. [S08, S09, S16–S18]

Supervisor and Lab Assessment

Dr Nóvoa is the named supervisor and appears in Imperial’s Aeronautics research context. The group’s 2022–2025 papers form a coherent thread from low-order model bias correction through thermoacoustic digital twins to sparse-data forecasting and assimilation-informed learning. That is strong evidence of methodological fit and recent activity. It does not establish the exact system, division of supervision, data access or high-dimensional work packages for AE0095. Ask how the PhD will connect to the current research programme and who will guide physical-system selection and computational scaling. [S15–S23]

Funding, Application and London Costs

The official page promises £23,805 tax-free per year and full tuition fees for Home, EU and international students over 3.5 years. Confirm award terms and any stipend review for 2027/28 entry. Imperial’s 2026 average London student living-cost estimate is £1,821–£1,893 per month, with lower and upper figures of £1,264 and £2,118. Costs include accommodation and daily expenses, exclude tuition and vary by housing. These are student planning figures, not an employee budget. The annual stipend is roughly £1,984/month when divided by 12; actual affordability depends on housing and personal circumstances. [S01, S06, S07]

The application is through Imperial’s online system with CV, personal statement, transcripts and details for two referees. Select “Aeronautics Research (PhD),” research group “Aero,” topic AE0095, supervisor Dr Andrea Nóvoa and Autumn 2027/8 or Spring 2027/8. Applicants can submit at most two Imperial PhD applications in the academic year and must choose different entry terms for two projects. References must arrive by 23:59 on 5 February 2027 UK time. General department guidance says advertised studentships do not require a separate research proposal; follow the current project checklist. [S01, S02]

Applicant Fit and Preparation

Formal requirements and preferred background

Formal eligibility is a first-class MEng/MSci or higher degree, or international equivalent, in a mathematical or engineering discipline; the page names Aeronautics, Mechanical Engineering, Mathematics, Statistics and Physics. Background in computational physics/mathematics, data assimilation or dynamical systems for fluid mechanics is advantageous, not mandatory. The project seeks a curious candidate comfortable at the interface of physics, mathematics and computation.

Prepare evidence of mathematical modelling, computation, state estimation, uncertainty quantification, dynamical systems or physical ML. A concise example could explain an imperfect model, the data available, the method used to update or test it, and the limits of the result. This is suggested preparation, not an extra formal screening rule.

Career Context, Risks and Next Steps

EngineeringUK estimates 6.3 million workers in UK engineering and technology. HESA reports 87% of 2023/24 graduates in work or further study 15 months after leaving. These broad outcomes do not predict PhD employment in data assimilation. Transferable skills include Bayesian methods, scientific machine learning, dynamical systems, numerical simulation and HPC; possible paths include engineering R&D, aerospace/energy, research software, data science, government labs and academia, with no guaranteed placement. [S11, S12]

Before applying

• Confirm the application fields, deadline, two references and start-term selection.
• Ask which low-order prototype and later multiphysics systems are currently in scope.

At interview

• Clarify how model bias, state uncertainty and parameter uncertainty will be separated and assessed.
• Ask what measurements or datasets are available, what “without violating physics” means in the chosen method, and how physical consistency will be tested.
• Confirm the planned progression from low-order to high-dimensional models, computing requirements and supervision arrangements.

Before accepting

• Confirm the funding package, tuition coverage, stipend, start date and duration in writing.
• Ask for the planned benchmark systems, data permissions, HPC/GPU route and concrete evaluation criteria.

Evidence and Sources

Official project, admissions and programme

• S01 PhD Studentship in Aeronautics: Real-time data assimilation for multi-physics flows (AE0095)
• S02 Aeronautics PhD application process
• S03 PhD study in Aeronautics

Research environment, supervisors and facilities

• S04 Aeronautics research groups
• S05 Aeronautics research facilities
• S08 Research Computing Service
• S09 Aeronautics computing resources
• S10 Aeronautics research and REF 2021 context
• S15 Aeronautics academic staff — Andrea Nóvoa
• S16 Real-time data assimilation research

Recent papers and field evidence

• S17 Scientific machine learning and real-time computation publications
• S18 Data-Driven Engineering publications
• S19 Real-time thermoacoustic data assimilation
• S20 A real-time digital twin of azimuthal thermoacoustic instabilities
• S21 Data-driven computation of adjoint sensitivities without adjoint solvers: An application to thermoacoustics
• S22 Data-assimilated model-informed reinforcement learning
• S23 Real-time forecasting of chaotic dynamics from sparse data and autoencoders

Funding, costs and career context

• S06 Cost of living in London (2026 student estimates)
• S07 Imperial postgraduate tuition fees 2026–27
• S11 EngineeringUK engineering and technology workforce briefing
• S12 HESA Graduate Outcomes 2023/24 summary
• S13 Early Career Researcher Institute introduction to HPC
• S14 South Kensington campus location