Master Thesis defense by Florian Sickert

Title: SR-LPNE - Developing a Benchmark for a Palaeoclimatological Neural Emulator

Abstract: The dynamics of Pleistocene glacial cycles have been an ongoing open question in palaeoclimatological research. While the influence of external insolation forcing on the climate system, the specific external feedback mechanisms, that caused a shift in dominant glacial cycles throughout the Pleistocene is still debated, with a variety of models reproducing the spectral powers of the records. These models range from dynamical systemmodels, forced numerical to ice-sheetmodels. We present SR-LPNE, the Sobolev-Rollout Late Pleistocene Neural Emulator, a neural network based surrogate model, that reconstructs the state space manifold of a nonlinear, dynamical system baseline for the dominant 100 kyr cycle during the Late Pleistocene. Through an iterative development process, we illustrate systematically common failure-modes in neural surrogate models such as the identity mapping or exposure bias. Using a Sobolev loss function and multi-step rollout, the model effectivly learns the underlying vector field of our baseline system.
Our results demonstrate, that this architecture can reconstruct the invariant measure of the Late Pleistocene glacial cycles. This provides a framework how to adapt neural networks as surrogate models for non-linear, long-term climate emulation, while mitigating the computational cost of full resolution climate simulations.

 Advisor: Prof. Dr. Markus Jochum