Supporting policymakers to incentivise effective and equitable land use change for Net Zero with AI
Background
British law has enshrined specific goals to meet Net Zero, reverse biodiversity loss, improve water quality and to ensure food and energy security in the coming decades. To meet legally binding targets, we must use large amounts of UK land for tree-planting, agroforestry, re-wilding, peatland restoration, bioenergy crops and other green technologies. But, even if we knew where and when to deploy each change to the land, most land is privately owned, in large part by farmers. Policymakers must develop payment schemes that deliver the required land use change whilst providing value for money for the taxpayer. To support policymakers we aim to develop real-time uncertainty-enabled prediction of the response to changes in the land across all relevant ecosystem services and risks, scalable agent-based models of uptake of agri-environment schemes and the tools to explore and compare the efficacy of potential policies.
PhD Opportunity
Environmental impact models in our decision support tools include crop, livestock and whole-farm models, tree and soil-carbon simulators, water and nutrient flow models, abundance for 1100 species and an agroforestry model. Run time varies from seconds to 4 hours to simulate the impacts of a single land use change into the future on architectures on a spectrum from a humble laptop to the JASMIN supercomputer. We deploy deep Gaussian processes (DGP) to capture these models and their uncertainty within our decision support tools. GPs are a standard workhorse for uncertainty quantification in models, and are part of the core CDT training. However, both they and their deep cousins cannot fit in large dimensional input spaces (more than 12 is a struggle) because of the curse of dimensionality. Yet the impacts we are concerned with all require daily weather forcings to run. Our crop model, for example, requires 9 daily time-series inputs for simulations to take us out to Net Zero by 2050 (~82000 parameters).
Searching for low dimensional representations of parameter spaces so that effective machine learning models (like DGPs) are effective has been called “”active subspace”” selection, “”dimension reduction”” and “”embedding””. This PhD will study the problem from a new perspective. Inspired by the success of the transformer in large language models, we will explore tokenised weather series and develop transformers for dimension reduction in climate impacts models. We will tailor these to GPs and DGPs in the first instance and compare their performance to existing approaches in the UQ literature such as Automatic Relevance Determination and active subspaces. Improvements to our existing emulation will quickly be deployed into our decision support systems if we can establish trust in their uncertainty quantification. We will work with policymakers in establishing effective trust in the decision support tools generated with the technologies developed in this PhD.
Applicant Profile
Students with a strong background in machine learning, statistics or mathematics who want to learn about and work with novel techniques in AI but who want to make a difference in the environment or in UK policy at the same time will be well suited to this PhD.
Other information
https://netzeroplus.ac.uk/project/add-trees/
https://www.exeter.ac.uk/research/leep/researchimpact/
