Uncertainty quantification on multiscale computer simulations for urban planning

Lead Research Organisation: University of Cambridge
Department Name: Engineering

Abstract

Urbanisation, population and economic growth have given rise to many large-scale complex challenges in urban environments. Understanding the state and evolution of urban environments poses a challenge for urban planners. Although conducting field experiments can provide insights into the dynamics of physical systems, they tend to be costly, time-consuming and not scalable. On the other hand, computer models representations of these systems (simulations) unlock opportunities for in silico experiments. Virtual experiments are faster, cheaper and safer than field studies. However, without field data, simulations do not know which reality to imitate. Therefore, a more realistic approach, known as data assimilation, involves leveraging available field data to calibrate computer model simulations.
Computer simulations are available at fine resolutions (microscopic models or equivalently agent-based (AB) models), or at a coarser level (macroscopic models). Examples of macroscopic models include ordinary or partial differential equations. Another class of simulators is the multiscale models which are a hybrid of microscopic and macroscopic models. The problem of calibrating these three classes of models is known as the inverse problem. Providing robust solutions to the inverse problem is vital for prediction, optimisation and control.
However, a number of challenges emerge. The continuous evolution of urban environments generates non-stationary, stochastic and volatile conditions that impede deterministic modelling. As a result, these sources of uncertainty need to be quantified. Although this work has been applied to computer simulations in biology, physics and environmental modelling, there haven't been many studies in urban planning.
Computer models are often computationally challenging due to the high dimensionality of parameter and output spaces or inefficiency of numerical solvers. To alleviate the computational burden, a class of surrogate models are employed. These models emulate simulators at much lower computational cost via approximation. Gaussian process (GP) emulation has been predominantly used to address heteroskedasticity and non-stationarity of model outputs. Although parallelised versions of GPs have been proposed, no work has attempted to induce sparsity in computer model outputs and achieve computational savings.
A related challenge involves efficiently calibrating simulators with high dimensional input and/or output spaces. Simulations can be computationally intractable even if only a small number of them is run, which motivates the need for simplifying them. Traditional approaches focus on employing feature engineering methods to reduce input/output dimensions as a pre-calibration step, but very few focus on exploiting the underlying latent structure of inputs/outputs.
The ultimate challenge is to combine microscopic and macroscopic simulators to perform multiscale inference. This approach involves fusing computational experiments available at multiple spatial and temporal resolutions of varying fidelity and dimension often computed at different input domains. This work also deals with model evaluation and selection. Although ad-hoc approaches have been developed to perform multiscale inference, there is no unified framework of doing multiscale computer model inference.
The objectives of the PhD work are to:
1. Advance the current state-of-the-art GP emulators for computationally expensive high-dimensional (in input and output space) computer model simulations.
2. Improve the efficacy of microscopic agent-based models and macroscopic equation-based models as well as the efficiency of their calibration and optimisation methods.
3. Develop a robust Bayesian multiscale framework that facilitates computer experimentation at varying levels of spatio-temporal resolution, fidelity and dimension.
4. Apply the developed methods to problems in London's transportation and air pollution management systems.

Publications

10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S02302X/1 01/10/2019 31/03/2028
2277256 Studentship EP/S02302X/1 01/10/2019 30/09/2023 Ioannis Zachos