Physics-Based Machine Learning: Opportunities and Challenges for Mantle Convection

Zenodo (CERN European Organization for Nuclear Research)

Abstract

This dataset contains the data and model outputs associated with the study on physics-based machine learning for mantle convection surrogate modeling. Specifically, it includes temperature distribution snapshots generated using ASPECT (v2.4.0), an open-source geodynamic modeling software, for 110 forward simulations of mantle convection in a 2-D quarter-shell geometry representing the Earth's mantle. The simulations span a viscosity range of 10²²–10²³ Pa·s and cover a modeled time period of one billion years. The dataset comprises 100 Latin Hypercube Sampling (LHS) training realizations and 10 randomly sampled validation realizations, used to construct and evaluate a Non-Intrusive Reduced Basis (NI-RB) surrogate model. The surrogate model combines Proper Orthogonal Decomposition (POD) with feedforward neural networks to enable rapid multi-query analyses of mantle temperature distributions, including global sensitivity studies and uncertainty quantification.

Authors 5

1. Marilina Valatsou

   ETH Zurich

   Affiliation as printed

   ETH Zurich

2. Denise Degen

   Technische Universität Darmstadt · GFZ Helmholtz Centre for Geosciences

   Affiliation as printed

   Technische Universität Darmstadt

   GFZ Helmholtz Centre for Geosciences

3. Juliane Dannberg

   GEOMAR Helmholtz Centre for Ocean Research Kiel

   Affiliation as printed

   GEOMAR Helmholtz Centre for Ocean Research Kiel

4. René Gassmöller

   GEOMAR Helmholtz Centre for Ocean Research Kiel

   Affiliation as printed

   GEOMAR Helmholtz Centre for Ocean Research Kiel

5. Florian Wellmann Aachen

   RWTH Aachen University

   Affiliation as printed

   RWTH Aachen University