Climate predictions of feels-like temperatures with online machine learning

Climate models provide crucial information for climate change mitigation and adaptation, but their predictions also lack necessary accuracy with respect to many atmospheric variables. Correctly predicting the probability of heat waves and especially the risk of associated heat illnesses, including heat strokes, has a great importance for society. The July 2022 European heat wave caused a UK national emergency with further socio-economic impacts due to the following drought. However preparation is possible as, for example, major outdoor sport events have adapted schedules to reduce the risk of heat illnesses.
Conventional atmospheric models predicting heat waves are built on physical laws for resolved dynamics but include so-called parameterizations of processes that are not explicitly resolvable, such as radiation, precipitation and surface fluxes. Decades of global observational data are available which machine learning can learn from and, if embedded inside an atmospheric model, could automatically correct the simulated climate.
With this NERC Independent Research Fellowship, I want to build a hybrid physics and data-driven atmospheric model that learns automatically from global data of feels-like temperatures to quantify the heat stress on the human body. I will develop so-called online learning for data-driven parameterizations in the atmospheric general circulation model SpeedyWeather.jl that I wrote over the last year. If the simulated precipitation learned to rain more over mountains, then the current offline learning methods would not increase soil moisture, nor impacting surface humidity and temperature. The proposed online learning however will provide physical consistency with machine learning in a climate model.
The main objective is to create a global dataset of feels-like temperature probabilities for the 2050s. The public and policymakers will have access to a catalogue quantifying the risk of critical heat exposure in their global location. This catalogue is like a weather forecast but for climate, for every season and time of day, providing vital information to mitigate and adapt to the expected human heat stress. With project partners in the UK and Germany, I will build on top of previous work on heat forecasts for the Tokyo 2020 Olympics.
The major scientific advance is the use of explainable machine learning inside a physics-based atmospheric model. With project partners in the USA, I will develop a hybrid climate model that can reveal missing physics in current modelling efforts by learning corrections to existing parameterizations. Only this combination of a model obeying physical laws with online learning for unresolved processes will allow climate simulations to be consistently enhanced with machine learning. This methodology will generally allow climate science to better understand the missing physics in current climate models.
The proposed hybrid climate models which include machine-learned physics does not exist yet, but this project will develop a prototype. Only with online learning can we combine physics knowledge and observations towards a new generation of reliable climate predictions. For individuals and policymakers, it is necessary to translate such predictions into crucial information. To better prepare against future heat waves, we have to predict and communicate effectively the feels-like temperatures and associated risks of heat illnesses.