Collaborative Research: CAIG: Artificial Intelligence-native embeddings for organic geochemistry

Petroleum contains molecular fossils: chemical traces of ancient organisms that can be preserved in rocks and oils for hundreds of millions of years. These molecules record information about the evolution of life and Earth's environment. But much of this record is difficult to interpret and depends on expert comparison of molecular ratios, one sample at a time. This project will build Artificial Intelligence (AI) tools that make molecular records searchable and comparable at scale. By linking thousands of characterized samples and interrogating all of them, scientists will be able to ask important questions about Earth's history, such as how biological communities changed across mass extinctions and how the molecular signatures of flowering plants spread through time and space. These tools also have practical uses for energy exploration, environmental protection, forensic oil-spill analysis, and for identifying petroleum origin. Broader impacts include training graduate and undergraduate students in geoscience and artificial intelligence, developing public educational materials, and releasing open-source software and datasets resulting from the project. At the core of the project is the learning of AI-native embeddings for organic geochemistry. These embeddings are numerical representations that place geochemically similar samples near one another in high-dimensional space. AI models will be trained to integrate biomarker measurements, bulk and isotopic data, chromatographic signals, and geological context. The models will also be trained to characterize uncertainty in sample age and paleogeographic position. A central test is whether the embeddings learned by these models improve similarity search, oil-family classification, source-rock prediction, and inference for partially characterized samples, measured against traditional geochemical and machine-learning baselines. The work also addresses two hypotheses about Earth's history: whether embeddings resolve extinction-associated shifts in biomarker composition, and whether they capture the spread of angiosperm molecular signatures across depositional environments and paleogeographic regions. 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.