deeP redUced oRder predIctive Fluid dYnamics model (PURIFY)

Simulating large area urban turbulence flows in real time accurately remains a pressing challenge, yet to be addressed. Reduce Order Modelling (ROM) provides a means of real time simulation; traditional model reduction methods such as balanced truncation, the reduced-basis method, and (balanced) proper orthogonal decomposition (POD), have been developed. However challenges remain: the traditional methods restrict the state to evolve in a linear subspace; in addition, big area simulation cannot be conducted in real time. In addition, it is hard to define how accuracy the derived ROM is since there is no appropriate error estimator derived for the ROM. Also, existing ROM is either physically or data-driven and there is no physically informed data-driven ROM. Advancing in machine learning and in observational and simulation capabilities offer an opportunity to integrate simulation and data science approaches more intensively.
The proposed research offers scientific advances within the field of fluid flow modelling in order to underpin the establishment of novel sophisticated tools that will allow real-time simulations and prediction of air flows (and subsequently personal exposure to air pollutants); principal features include real-time air pollution predictions for next few hours. This can help minimise exposure to the population, especially vulnerable young/elderly patients. Individuals will be able to decide whether to go outside or not, exercise or make their daily plans.
The proposed new model combines physical models and data sciences technologies, in particular, the deep learning for predicting the non-linearity of turbulence flows, thus making the air flows predictions more reliable and accurate. An autoencoder deep neural network will potentially capture the non-linearity of urban turbulence flows and more accurate than the traditional model reduction methods (e.g. POD). Commercialisation of the research outputs will be undertaken in partnership with VortexIoT, Looker Tech and Spire Global. These are international leading companies in the fields of sensor technology, software company and provision of technical professional services.
The approach is based on an advanced fast-running computational model to manage and predict the airflow, air quality in a city, guide effective responses in emergencies and help people reduce the air pollution exposure time. The development of novel deep learning ROM, reducing computational times by several orders of magnitude, will make currently unsurmountable problems tractable:- e.g. detailed air flows through a big area or an entire city. The specific technology that distinguishes this project are the potential use of deep learning reduced order modelling, new computational domain decomposition, new error analysis for DLROM and data assimilation method.
The research objectives are: to develop deep learning based ROM framework with domain decomposition methods; to develop ROM framework with Autoencoder network that will project the system into a non-linear subspace, thus increasing accuracy; to develop ROMs with variable material properties such as variable initial or boundary conditions (different wind direction for example); ROM based data assimilation and optimisation methods; to perform an error analysis and optimal improvement for ROMs using machine learning methods.
The research has substantive health, environmental and economic impacts. Beneficiaries may include the general public through fast response to avoid exposure to air pollution technical professional services consultancy companies, local and national government environment bodies (e.g. NRW in Wales), computational engineering companies who will benefit from more efficient fluid flow model outcomes. In addition, public sector (spend budget savings in health) and air quality sensor device manufacturers incorporating the proposed modelling approach to enhance their offering (increased profits from sales).

The rapidly increasing global population and occupation of urban areas has exacerbated major societal challenges: poor air quality; insufficient water availability/quality; waste disposal issues; energy consumption. These issues highlight the demand for innovative strategies and methods to design and manage cities. Key to this goal is effective decision-making tools, underpinning strong city planning. Existing systems (Urban Flows Observatory; Urban Observatory) gather data from mobile and fixed sensors and deploy sophisticated visualisation tools. Simulation tools to model urban flows (Street-in-Grid model) illustrate the concentrations of air pollutants in complex urban canopy configurations and the background concentrations in the overlying atmosphere. Climate models studying interactions of atmosphere, oceans, land surface and ice have multiple linkages to air quality as many air pollutant sources also emit CO2 and greenhouse gases. However, existing traditional models notably lack accuracy in predicting non-linear systems in large computational areas in real time (e.g. urban turbulence flow). This tempers the ability to plan and protect (especially for the elderly, infirm or young) against exposure to low air quality.
The proposed research will innovatively surmount these challenges to facilitate the development of a complete efficient modelling system, combining physical models and data sciences technologies, in particular, the deep learning for predicting the non-linearity of turbulence flows, thus making the air flows predictions fast (i.e. real time), more reliable and accurate. This will involve setting up an advanced fast-running computational model to manage and predict the airflow, air quality in a city, and guide effective response in emergencies and help people reduce the air pollution exposure time.
The eventually commercialised product will orchestrate salient socio-economic impact. The impact beneficiaries and categories are:

Health Impact - A new fast response model will orchestrate marked benefits for people living in urban centres. improved public health outcomes for vulnerable people, via more informed air quality forecasts, as data analysis tools will be incorporated to modern (e.g. phone apps), and traditional media channels. This will also enable guide response in emergencies, supporting the delivery of healthcare professional services by reducing bottlenecks from hospital and health centre admissions.

Knowledge - Beneficiaries include technical professional services/consultancy companies, local and national government environment bodies (e.g. NRW in Wales), computational engineering companies. These will benefit from knowledge enhancement/uptake of new practises and technology within industry.

Economic Impact - The development and eventual commercialisation of the model will realise significant financial benefits for the industrial partners (and supply chains). No competitive solution exists with the capacity to manage and accurately predict airflow and report air quality real-time variation in a city environment. Consultancy companies could offer superior services to their clients translating in higher revenues. Air quality monitoring hardware developers will be able to offer an enhanced service coupling their real time data collection with our powerful prediction tool.
Finally, local governments and the UK economy can be strengthened from innovation and cost-effective technology to underpin strong city planning

Environment Impact - Energy efficient building solutions and low carbon transport systems will reduce air pollution from travel and traffic.

Public Engagement -The use of Computational Engineering to address the challenges society is facing with urbanisation promises to stimulate public interest and lead to heightened understanding. Beneficiaries include school children; respiratory patient support groups; environment regulator and policy makers.