Explainable Artificial Intelligence in Meteorology and Climate Science: Model Fine-Tuning, Calibrating Trust and Learning New Science

Abstract

In recent years, artificial intelligence and specifically artificial neural networks (NNs) have shown great success in solving complex, nonlinear problems in earth sciences. Despite their success, the strategies upon which NNs make decisions are hard to decipher, which prevents scientists from interpreting and building trust in the NN predictions; a highly desired and necessary condition for the further use and exploitation of NNs’ potential. Thus, a variety of methods have been recently introduced with the aim of attributing the NN predictions to specific features in the input space and explaining their strategy. The so-called eXplainable Artificial Intelligence (XAI) is already seeing great application in a plethora of fields, offering promising results and insights about the decision strategies of NNs. Here, we provide an overview of the most recent work from our group, applying XAI to meteorology and climate science. Specifically, we present results from satellite applications that include weather phenomena identification and image to image translation, applications to climate prediction at subseasonal to decadal timescales, and detection of forced climatic changes and anthropogenic footprint. We also summarize a recently introduced synthetic benchmark dataset that can be used to improve our understanding of different XAI methods and introduce objectivity into the assessment of their fidelity. With this overview, we aim to illustrate how gaining accurate insights about the NN decision strategy can help climate scientists and meteorologists improve practices in fine-tuning model architectures, calibrating trust in climate and weather prediction and attribution, and learning new science.

Predicting Slowdowns in Decadal Climate Warming Trends With Explainable Neural Networks

“…We implement the LRP z rule for back propagation, which was found by Mamalakis et al. (2021) to be a well performing XAI method using a benchmark climate data set similar to ours. To improve interpretation and reduce the amount of noise in the LRP heatmaps, we only focus on positive areas of relevance, which are features that contribute positively to the ANN's prediction output.…”

Carefully Choose the Baseline: Lessons Learned from Applying XAI Attribution Methods for Regression Tasks in Geoscience

“…Appropriate use and/or experimentation with multiple baselines will be advantageous for many XAI-pursued goals in geoscientific applications. These include better deciphering the decision-making process of the network (e.g., McGovern et al, 2019;Toms et al, 2020), accelerating learning new science (e.g., Sonnewald and Lguensat, 2021;Clare et al, 2022;Mamalakis et al, 2022a) and potentially helping identify problems in the training dataset (Sun et al, 2022).…”

“…EXplainable Artificial Intelligence (XAI) aims to provide insights about the decision-making process of AI models and has been increasingly applied to the geosciences (e.g., Toms et al, 2021;Ebert-Uphoff and Hilburn, 2020;Hilburn et al, 2021;Barnes et al, 2019Barnes et al, , 2020Mayer and Barnes, 2021;Keys et al, 2021;Sonnewald and Lguensat, 2021). XAI methods show promising results in calibrating model trust, and assisting in learning new science (see for example, McGovern et al, 2019;Toms et al, 2020;Sonnewald and Lguensat, 2021;Clare et al, 2022;Mamalakis et al, 2022a). A popular subcategory of XAI is the so-called local attribution methods, which compute the attribution of a model's prediction to the input variables (also referred to as "input features").…”

Updates on Model Hierarchies for Understanding and Simulating the Climate System: A Focus on Data‐Informed Methods and Climate Change Impacts

“…Efforts like XAI have demonstrated the potential to uncover new statistical relationships between model inputs and outputs. In addition to improving existing climate models, and complementing physics‐based models in the hierarchy through better parameterizations, calibration, and UQ, some claim that ML‐based models may also provide a novel approach to modeling physical systems and unraveling new climate patterns, teleconnections, and mechanisms through a potentially more nuanced process isolation and learning (Mamalakis et al., 2022)—which is one of the central goals of having a model hierarchy. Incorporation of ML algorithms has already led to novel discoveries in medicine (Stokes et al., 2020), astronomy (Valizadegan et al., 2022), chemistry (Hueffel et al., 2021) and mathematics (Davies et al., 2021).…”

“…In practice, the first approach to interpret an ML model is to manually verify the correlations between outputs are reasonable (e.g., Rasp & Thuerey, 2021). Following this initial step, there exists a wide variety of more sophisticated methods that can be used to look inside the “black box” of ML models and explain the information hotspots that establish the relationships learned (Mamalakis et al., 2022; McGovern et al., 2019). These methods are often labeled as “explainable AI” (XAI), which can be defined as:…”