A NON-DETERMINISTIC DEEP LEARNING BASED SURROGATE FOR ICE SHEET MODELING

• Main Author: Jordan, Hannah
• Format: Thesis
• Language: unknown
• Published: University of Montana 2022
• Subjects: glaciology; surrogate; statistical; machine learning; autoencoder; deep learning; Numerical Analysis and Scientific Computing; Statistical Models; Ice Sheet
• Online Access: https://scholarworks.umt.edu/etd/11991; https://scholarworks.umt.edu/context/etd/article/13102/viewcontent/Jordan_Hannah_Thesis.pdf

Surrogate modeling is a new and expanding field in the world of deep learning, providing a computationally inexpensive way to approximate results from computationally demanding high-fidelity simulations. Ice sheet modeling is one of these computationally expensive models, the model used in this study currently requires between 10 and 20 minutes to complete one simulation. While this process is adequate for certain applications, the ability to use sampling approaches to perform statistical inference becomes infeasible. This issue can be overcome by using a surrogate model to approximate the ice sheet model, bringing the time to produce output down to a tenth of a second or less. In this paper, we introduce the use of a conditional variational autoencoder as a surrogate model for approximating an ice sheet model producing surface velocity predictions. For a fair comparison, we test both deterministic and stochastic approaches and discuss the drawbacks and benefits to both model types. We train a standard vanilla neural network architecture, a neural network architecture using dropout and normalization, and a neural network with added dimensionality reduction using principal component analysis. These surrogate models produce output that is representative of the high-fidelity data, but there is variability between the surrogate and high-fidelity model. This divergence cannot be determined for a deterministic model without further analysis such as model ensembling. The use of a stochastic network, such as the conditional variational autoencoder, provides a solution to this problem. This network provides us with a method to quantify the uncertainty within the surrogate using the model's natural stochasticity. This implementation has the potential to be applied across multiple fields because of the black box nature of the architecture.