Combining Tectonics and Machine Learning to Constrain Plate Reconstruction Models Through Time

ended

Plate tectonics processes continuously destroy oceanic crust, which contain the most reliable record of plate motion. There is therefore little data to constrain net rotation of the lithosphere with respect to the deep mantle, constraints on which are required to produce accurate reference frames for plate motion, The location of intra-oceanic plate boundaries and bathymetry in the geological past are also lost. I will use state-of-the-art numerical convection simulations combined with state-of-the-art machine learning techniques to put constraints on both net rotation and the location of plate boundaries with uncertainty estimates. This is possible due to the self-organising and statistically predictable nature of plate tectonics. I will develop one set of neural networks to make inferences for net rotation with uncertainties given observation of continent positions and movement. The networks will take both synthetic and real geological observations as training inputs and produce estimates for net rotation. They will be thoroughly tested using synthetic data and benchmarked using present-day Earth data, thereby testing both the networks and the physics behind the convection simulations. The networks will then be applied to the geological past. A second set of networks will treat the lack of information on oceanic plate boundaries as an image completion problem to fill the gaps in geological data. They will be trained to produce proposals for the location and type of oceanic plate boundaries that are consistent with the physics behind tectonic motion and mantle convection. The networks learn about the physics from the database of convection simulations. These proposals can be assessed against geological and palaeo-oceanographic data, provide suggestions for alternative solutions, give an indication of uncertainties and guide future data collection and modelling work.