Climate change over the 21st Century is expected to affect many aspects of the earth system.
Projections of these effects are often considered separately for different quantities, but it is the
combined effect of these changes which is most relevant to understanding their impacts.
Identifying the combined changes across climate variables, however, is not entirely
straightforward given that each variable’s response may exhibit different evolutions in time and
different spatial patterns – and it is the combined patterns of change projected by climate models
that are most relevant for impacts. Finally, even if these patterns can be identified, there is large
uncertainty for when we will be able to detect them in the observations due to climate noise.
Thus, any attempt to understand how these patterns of combined change will amplify or vary
over the 21st Century requires explicit consideration of the signal-to-noise ratio.
Motivated by recent advances in using artificial intelligence to autonomously detect complex
patterns in many different settings, we propose a truly novel method based on artificial neural
networks to detect 21st Century patterns of combined change and quantify when their signal will
emerge from the background climate noise. This will be done under the umbrella of three main
goals: (1) Develop a state-of-the-art neural network architecture to detect forced time-varying
combined patterns of change of impact-related earth system quantities, (2) Identify the combined
patterns of change over the 21st Century, determine how they change over time, and quantify
when these patterns will emerge from the background of climate noise within CMIP6, and (3)
Quantify the extent to which combined patterns of change have already emerged in observations.
The PIs have already developed a successful prototype of the neural network architecture for a
single climate variable (e.g. temperature) and applied the trained network to observations, and so
the extension to combinations of climate quantities is at the center of this proposal. The proposed
work will focus on combinations of climate quantities (including National Climate Assessment
indicators) that lead to impacts related to weather extremes, wildfire occurrence, drought and
poor air quality. Furthermore, we fully expect the methodology to be applicable across a wide
range of variables and impacts.
Relevance and Suitability for NOAA
The project addresses the NOAA MAPP 21st Century Integrated US Climate Predictions and
Projections competition by developing and applying the innovative methodology of neural
networks, a type of machine learning (ML) method. ML methods are specifically highlighted in
the call. The outcomes of this method will address Priority Areas A and C: Priority A will be
addressed by identifying the patterns of combined change over the 21st Century, determining
how they change over time, and quantifying when these patterns will emerge from climate noise.
Priority C will be addressed by applying the trained neural network to the observations and a
combination of NCA indicators to quantify how these combined effects have changed in recent
decades. PI Barnes has extensive experience studying atmospheric responses to climate change.
Co-PI Anderson is an expert in machine learning and developing neural network architecture and
Co-PI Ebert-Uphoff is an expert in complex networks and the intersection of climate science and
artificial intelligence. All three have worked together on a successful prototype.