Project: Improving the representation of uncertainty in climate-change prediction
Supervisors: Etienne Roesch (UoR), Ted Shepherd (UoR) and Marilena Oltmanns (NOC)
Project Description:
Statistical climate models that rely on past observations to predict future changes have started to yield divergent results, when comparing expected and observed annual mean temperatures (Schmidt, 2024). Similar discrepancies exist for long-term observed trends in many regional aspects of climate, compared with the predictions from physical climate models (Shaw et al., 2024). This may in part be due to changing conditions, resulting in a growing “knowledge gap” in how the climate system operates. But it may also be due to fundamental issues with the statistical inference methods and the way they treat the uncertainty arising from the computational methods employed by modellers. It is therefore critical to examine the practice of statistics by climate scientists, to formulate recommendations that will strengthen trust in climate predictions and projections.
Bayesian inference is a versatile method that can incorporate heterogeneous sources of uncertainty, about measurements, model structure and parameterization, and computational constraints, such as trade-offs between resolution and complexity. Yet it isn’t routinely employed in climate-change science . Bayes can manage the large parameter spaces available to modellers, that need to be interpreted for models to be meaningful. It can also allow for the examination of the decision making workflow itself, highlighting steps that can be more sensitive to issues and may add noise to estimates and predictions.
The project combines expertise in climate physics with state-of-the-art mathematical approaches to evaluate current statistical practices in climate-change science, identify potential issues, and develop a Bayesian framework to quantify uncertainty in climate-change predictions and projections.
