PREDICTING INTERGLACIAL TRANSITIONS WITH MACHINE LEARNING AND THE MILANKOVITCH CYCLES

Milankovitch cycles have long been theorized to drive the glacial-interglacial pattern which has predominated throughout the Pleistocene period beginning 2.58 million years ago. However modeling efforts to predict temperature with these cycles have struggled to explain some observed features such as marine isotope stages (MIS) 5 and 11. The present project used machine learning and artificial intelligence to create generalizable models which can translate Milankovitch cycle data into specific temperature predictions.

Machine learning proved too capable at predicting reconstructed temperatures (r² = 0.88-0.95) - to the point that it could predict temperatures observed in ice core data from arbitrarily chosen sine waves (r² = 0.74). A number of robustness tests and procedures are proposed to address the use of machine learning in handling time series problems for periodicities which may be present in the geologic record. These include:

1. Both randomized and MIS stage cross-validation
2. Evaluating the structure of successful models
3. Counterfactual tests on simulated cyclical patterns

These three factors were key to identify whether a given model memorized a specific pattern or learned a generalizable lesson from Milankovitch cycles. These cross-validation and interpretability methods were more suited to validate AI applications compared to traditional hypothesis testing frameworks.