Artificial intelligence in the paleogeosciences: progress, challenges, and opportunities

Abstract:

“The past is the key to the future.” In a seminal paper from the 1980’s, Bruce Doe recognized the importance of the then emergent field of paleoclimatology to decisions federal agencies would have to make in the face of rising greenhouse gases concentration. The promise of paleoclimatology is two-folds: (1) to assess the current climate change in the context of climate variations across the decades and centuries, and (2) to validate climate model outputs. But it wasn’t until recently, with advances in geoinformatics, that paleoclimatology could fully fulfill all of its promises.

Advancements in artificial intelligence have allowed (paleogeo)scientists to make their workflows more efficient. Knowledge representation has helped with the day-to-day work of scientists by broadening access to data (e.g. Google Dataset, EarthCube’s Geocode) and to automated cutting-edge data analytic tools. Here, I will present results from the LinkedEarth project, which is dedicated to standardize paleoclimate datasets to make them more Findable, Accessible, Interoperable, and Reusable (FAIR). These standards were then used to create software libraries in R (GeoChronR) and Python (Pyleoclim) to analyze paleoclimate datasets without requiring excessive manual data transformation. Our latest project, autoTS, builds on these capabilities to create an intelligent system for the analysis of time series data. The system is based on expert-grade abstract workflows, which can be instantiated with methods appropriate for the data. Looking forward, natural language processing can help scientists annotate their datasets based on these emergent standards. Finally, advances in deep learning have opened doors for a new use of paleogeoscience data: weather forecasting. In this case, a machine can learn from the past to predict the future, fulfilling the promise that “the past is the key to the future”.