CAIG: Navigating the Climate Science Deluge: Training Language Models to Assist in Comprehensive Assessments

Reviewing and summarizing existing climate research results is critical for guiding future research and policy. The exponential growth of climate research literature makes working one paper at a time no longer efficient. The proposed work aims to discover new, more efficient ways to scour the scientific literature.

This project will use artificial intelligence to build a climate science "chat bot" assistant. This assistant will be trained specifically for climate science literature. It will have the ability to independently read and report peer-reviewed papers.

This will enable it to always include the latest peer-reviewed results published. It will help researchers review the literature more comprehensively and in less time. Improved assessment reports will contribute to better-informed climate policy.

Addressing the assessment bottleneck will accelerate climate research by better identifying the multiple lines of evidence needed to assess consensus and consistency. This project will also create fundamental advances in artificial intelligence through use-driven research in climate science. Current large language models often emphasize “chat bots” that address general topics.

This project will develop new methods that can teach large language models specialized knowledge and skills without human intervention through large-scale data annotation efforts. These methods include synthetic training-data generation and interacting large language models. To improve huma interaction, the project will also develop methods for users to identify what types of inputs are difficult for the artificial intelligence to understand.

By combining these new methods into an artificial intelligence assistant, the project will initiate a new approach to conducting comprehensive assessments.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.