Abstract:The challenge that climate change poses to humanity has spurred a rapidly developing field of artificial intelligence research focused on climate change applications. The climate change AI (CCAI) community works on a diverse, challenging set of problems which often involve physics-constrained ML or heterogeneous spatiotemporal data. It would be desirable to use automated machine learning (AutoML) techniques to automatically find high-performing architectures and hyperparameters for a given dataset. In this work, we benchmark popular AutoML libraries on three high-leverage CCAI applications: climate modeling, wind power forecasting, and catalyst discovery. We find that out-of-the-box AutoML libraries currently fail to meaningfully surpass the performance of human-designed CCAI models. However, we also identify a few key weaknesses, which stem from the fact that most AutoML techniques are tailored to computer vision and NLP applications. For example, while dozens of search spaces have been designed for image and language data, none have been designed for spatiotemporal data. Addressing these key weaknesses can lead to the discovery of novel architectures that yield substantial performance gains across numerous CCAI applications. Therefore, we present a call to action to the AutoML community, since there are a number of concrete, promising directions for future work in the space of AutoML for CCAI. We release our code and a list of resources at <a class="link-external link-https" href="https://github.com/climate-change-automl/climate-change-automl" rel="external noopener nofollow">this https URL</a>.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are: **evaluate the performance of existing Automated Machine Learning (AutoML) techniques in these three high - impact climate change applications, namely climate modeling, wind power generation prediction, and catalyst discovery, and explore their limitations**. Specifically, the paper answers the following two research questions by benchmarking the performance of popular AutoML libraries on these three tasks: 1. **RQ 1**: Can current off - the - shelf AutoML techniques significantly improve performance in high - impact Climate Change AI (CCAI) applications, surpassing human - designed models? 2. **RQ 2**: If not, what are the key limitations and weaknesses of the existing techniques? ### Main Findings: - **Experimental Results**: Although AutoML has shown some improvement in certain tasks (such as ClimART), in most cases, AutoML has failed to significantly outperform human - designed models. In particular, in the tasks of wind power generation prediction and catalyst discovery, the performance of AutoML is even worse than that of human - designed models. - **Limitations Analysis**: The paper points out that the main weakness of current AutoML methods lies in the fact that they are mainly designed for computer vision (CV) and natural language processing (NLP) tasks, lacking search spaces for spatio - temporal data and Physics - constrained ML. For example, existing AutoML techniques do not specifically design search spaces for spatio - temporal data in climate change applications, nor do they consider how to combine physical constraints to optimize model architectures and hyperparameters. ### Call to Action: Based on these findings, the authors call on the AutoML community to focus on the following directions: 1. **Design search spaces suitable for spatio - temporal data and physically - constrained tasks**: This will help discover new architecture combinations, thereby achieving significant performance improvements in multiple climate change applications. 2. **Explore other AutoML sub - fields**: Besides hyperparameter optimization (HPO) and neural architecture search (NAS), data augmentation, pre - processing, and continuous monitoring and maintenance of deployed models can also be studied. 3. **Expand to more climate change applications**: Future work can test AutoML in other high - impact applications, such as model predictive control of buildings and optimal power flow. ### Conclusion: Although existing AutoML techniques have not yet shown significant advantages in climate change applications, by specifically improving search spaces and optimization strategies, AutoML is expected to make greater contributions to combating climate change in the future.

AutoML for Climate Change: A Call to Action

Tackling Climate Change with Machine Learning

Machine learning and artificial intelligence to aid climate change research and preparedness

Leveraging Artificial Intelligence for Climate Change Mitigation: Predictive Models, Optimization Techniques, and Sustainable Solutions

AI and machine learning in climate change research: A review of predictive models and environmental impact

Climate Intervention Analysis using AI Model Guided by Statistical Physics Principles

AI Foundation Models for Weather and Climate: Applications, Design, and Implementation

ClimateLearn: Benchmarking Machine Learning for Weather and Climate Modeling

Position: A Call to Action for a Human-Centered AutoML Paradigm

AI-enabled strategies for climate change adaptation: protecting communities, infrastructure, and businesses from the impacts of climate change

Climate-Invariant Machine Learning

Robustness of AI-based weather forecasts in a changing climate

Machine Learning for the Physics of Climate

AutoML: A survey of the state-of-the-art

Artificial Intelligence for Climate Change Biology: From Data Collection to Predictions

Machine learning for climate physics and simulations

Efficient Climate Simulation via Machine Learning Method

Extreme AutoML: Analysis of Classification, Regression, and NLP Performance

AI-Driven Carbon Emissions Tracking and Mitigation Model

Towards Green Automated Machine Learning: Status Quo and Future Directions