Unravelling interior evolution of terrestrial planets using Machine Learning Siddhant Agarwal, ESO AIA2019 PhD Supervisors Dr. Nicola Tosi Prof. Dr. Doris Breuer Dr. Pan Kessel Dr. Grégoire Montavon Prof. Dr. Klaus-Robert Müller Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine Learning DLR.de • Chart 1
37
Embed
Unravelling interior evolution of terrestrial planets using Machine … · Unravelling interior evolution of terrestrial planets using Machine Learning Siddhant Agarwal, ESO AIA2019
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Unravelling interior evolution of terrestrial planets using Machine Learning
Siddhant Agarwal, ESO AIA2019
PhD Supervisors
Dr. Nicola Tosi
Prof. Dr. Doris Breuer
Dr. Pan Kessel
Dr. Grégoire Montavon
Prof. Dr. Klaus-Robert Müller
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 1
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 2
• Introduction to mantle convection and the inverse problem
• Data used for inversion
• Results using Mixture Density Networks
• Next steps using this approach
• Acknowledgements
• References
Agenda
Introduction
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 3
Introduction
We are interested in understanding thermal evolution of terrestrial planets like Mars and Earth.
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 4
[1]
Introduction
We are interested in understanding thermal evolution of terrestrial planets like Mars and Earth.
Mantle convection is an important driver of it.
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 5
[2]
[1]
Introduction
Over geological time scales, rocks behave likes fluids.
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 6
[7]
Introduction
Over geological time scales, rocks behave likes fluids.
Hence we use fluid dynamics simulations to study mantle convection.
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 7
[7]
Introduction
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 8
Inputs
Viscosity
Initial temperature
Radiogenic elements
⁞
Outputs
Surface heat flux
Radial contraction
Crustal thickness
⁞
In-house C++ code
Conservation of mass:𝜕𝜌
𝜕𝑡+ ∇ . 𝜌𝑢 = 0
Conservation of momentum:
𝐷𝜌𝑢
𝐷𝑡= −∇𝑃 + ∇. 𝜏 + 𝜌 Ԧ𝑔
Conservation of energy:
𝜌𝑐𝑝𝐷𝑇
𝐷𝑡− 𝛼𝑇
𝐷𝑃
𝐷𝑡= ∇. 𝑘∇𝑇 + 𝜌𝐻 + 𝜙
• Mantle convection is governed by several poorly constrained parameters and initial conditions
Introduction
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 9
Inputs
Viscosity
Initial temperature
Radiogenic elements
⁞
Outputs
Surface heat flux
Radial contraction
Crustal thickness
⁞
• Mantle convection is governed by several poorly constrained parameters and initial conditions
• In planetary science, the outputs are observable (…sometimes)
Introduction
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 10
Inputs
Viscosity
Initial temperature
Radiogenic elements
⁞
Outputs
Surface heat flux
Radial contraction
Crustal thickness
⁞
• Mantle convection is governed by several poorly constrained parameters and initial conditions
• In planetary science, the outputs are observable (...sometimes)
• Need Machine Learning for rapid inversion in high-dimensional spaces; Monte Carlo methods are
computationally unfeasible
[3]
Dataset
Siddhant Agarwal • ESO AIA2019: Unravelling interior evolution of terrestrial planets using Machine LearningDLR.de • Chart 11
Dataset
• Generated some 3200 2D, quarter-cylinder evolution simulations for Mars, with: