Project institution: University of Glasgow
Project supervisor(s): Dr Raimondo Penta (University of Glasgow), Dr Vinny Davies (University of Glasgow), Prof Jessica Davies (Lancaster University), Dr Lawrence Bull (University of Glasgow) and Dr Matteo Icardi (University of Nottingham)
Overview and Background
While only covering 3% of the Earth’s surface, peatlands store >30% of terrestrial carbon and play a vital ecological role. Peatlands are, however, highly sensitive to climate change and human pressures, and therefore understanding and restoring them is crucial for climate action. Multiscale mathematical models can represent the complex microstructures and interactions that control peatland dynamics but are limited by their computational demands. GPU and Exascale computing advances offer a timely opportunity to unlock the potential benefits of mathematically-led peatland modelling approaches. By scaling these complex models to run on new architectures or by directly incorporating mathematical constraints into GPU-based deep learning approaches, scalable computing will to deliver transformative insights into peatland dynamics and their restoration, supporting global climate efforts.
Teaser Project 1: Scalable Mathematical Modelling of Peatlands
Objectives: This project will explore how we can do scalable modelling and inference on mathematical models of peatlands. The project will take existing microscale models for peatlands and look at how we can perform mathematical optimisation to the learn the complex parameters of the mathematical model. The focus will then be on looking at how the model can be upscaled and improved, focusing on computational inference methods that will be applicable as the model gets expanded and becomes more computationally infeasible.
Methods: The project will use scalable mathematical processing and optimisation techniques, looking at how they compared to computational statistical inference methods such a Bayesian optimisation. The peatland model will be used for simulations and analysed to understand how it can be improved to model the complex non-linear processes. Future work as part of a potential PhD project would involve extending model and adapting it to be able to run in high performance computing environments and extending the optimisation techniques to work in this scenario.
PhD Project: The main purpose of the PhD project will be scaling up the peatland models, adding more features and scaling them to be able to run across computer clusters and on GPUs with the eventual aim of extending this to Exascale computing. Advanced mathematical techniques will be used to upscale from micro- to macroscale models incorporate nonlinear instabilities such as wrinkling and surface patterning. Computational methods will then be extended to focus on predicting long-term peatland behaviour under restoration scenarios and climatic stressors. The integration of experimental data for validation and refinement of models will ensure practical applicability.
Teaser Project 2: Mathematically Informed Machine Learning for Scalable Peatland Modelling
Objectives: This project will explore how we can used emulation techniques for scalable parameter inference and uncertainty quantification in an existing model for peatlands. We will use parallelised computing to run multiple simulations of the peatland model and then use GPU based deep learning methods to build an emulator. The emulator will provide a computational cheaper version of the original model, allowing us to use Bayesian inference in previously computational infeasible scenarios giving the ability to estimate the model parameters and their associated statistical uncertainty.
Methods: The project will make use of deep learning architectures that are designed to be specifically scalable to GPUs and eventually Exascale type infrastructures. Specifically, the emulation methods will use and compare deep neural networks and deep Gaussian processes to link model parameters to observed model outcomes. Optimisation and Bayesian inference will then be carried out using the emulator within the context of an inverse problem. Future work as part of a potential PhD project would involve extending these methods into more complex deep learning frameworks, e.g. physics informed machine learning or graph neural networks.
PhD Project: A potential follow-on PhD project would focus on incorporating the mathematical models directly into the deep learning structures and linking the model to real data. This could be achieved by making the models more scalable by replacing the mathematical finite element methods via GPU trained deep learning alternatives or through methods from the physics informed machine learning literature. . Linking the model to real data will also be computationally challenging, building either directly on the emulation methods from the initial project or through the mathematical informed machine learning methods that have been developed. Essentially, this project will aim to link this model to real world applications that can help us gain a better understanding of the structure of peatlands.
References and Further Reading
