Predictability

Predictability is not just about making predictions, but more about understanding which aspects of the climate system can be predicted and why. Some changes, in some regions and on some timescales, are far more predictable than others. Knowing these limits is how we turn climate science into information people can actually use.

Several threads run through this work: identifying the large-scale connections that make some signals predictable, getting that information down to the local scales where decisions are made (while being honest about how much we can trust it), and breaking the climate system down to its essential, predictable parts.

*El Nino Southern Oscillation teleconnections. From climate.gov.*

Which signals are predictable?

Teleconnections describe the idea that a perturbation in one region of the globe can cause effects far away. A classic example is the El Niño Southern Oscillation, which occurs every 2-7 years and produces warmer sea surface temperatures in the central or eastern Pacific with far-reaching effects on weather worldwide. Because these effects are distant and result from complicated Earth system interactions, they are difficult to disentangle: which responses are due to a specific teleconnection, how strong are they, and how long do their signals take to appear? We build systematic ways of identifying teleconnections and the timescales on which they operate - using system identification, transfer functions, network science, and spread spectrum communication - so we can say more rigorously what is predictable.

*A schematic of how downscaling goes from large-scale to very local information. Reproduced from Khan and Pilz (2018) under a Creative Commons License.*

Getting to local scales

Even when a signal is predictable at large scales, people need information at local scales to make decisions. Downscaling turns the large-scale output of a global climate model into that local detail: statistical downscaling estimates it with mathematical tools, while dynamical downscaling feeds the large-scale information into a regional model. Producing a downscaled dataset and verifying its accuracy is hard, requiring bias-correction and extrapolation steps that are often specific to a location. Our work on downscaling is figuring out which methods work best for which situations, quantifying how much we can trust that local information, and showing added value - when might you want to use downscaling in the first place?

*Machine learning trained on climate models to provide data-driven climatological forecasts. From Weber et al. (2019).*

Breaking it down to essentials

Climate models are a central pillar of climate science, but they're also very complex and computationally expensive. Reduced-order models and emulators strip the system down to its essential parts, capturing the behavior that matters. This gives us better information about not only what is predictable, but why it's predictable.

Predictability

Predictability

Which signals are predictable?

Getting to local scales

Breaking it down to essentials

Kravitz Research Group resources