| Popis: |
At present, exploring the space of rheological parameters in models of glacial isostatic adjustment (GIA) and relative sea level (RSL) which incorporate laterally variable Earth structure is computationally expensive. A single simulation using the Seakon model (Latychev et al., 2005), using contemporary high-performance computing hardware, requires several wall-days & ≈ 1 core-year for one RSL simulation from late Marine Isotope Stage 3 to present day. However, it is well established that the impact from laterally variable mantle viscosity and lithospheric thickness on RSL and GIA is significant (Whitehouse, 2018). We present initial results from using the Tensorflow (Abadi et al.) framework to construct artificial neural networks that emulate the difference in the rate of change of relative sea level and relative radial displacement between model configurations using spherically symmetric (SS) and laterally variable (LV) Earth structures. Using this emulator we can accurately sample the parameter space (≈ 360 realisations of the background (SS) structure) for a given realization of lateral Earth structure (e.g. viscosity variations derived from shear-wave tomographic models) using ≈ 1/10th the amount of parameter vectors as a training set. Average misfits are O(0.1-1%) of the total RSL signal when using the emulator to adjust SS GIA model output to incorporate the impact from LV. We shall report on two case studies which allow us to examine the influence of lateral Earth structure on inferences of background (i.e. global-mean) viscosity. For these case studies, the emulator, in conjunction with a fast SS GIA/RSL model, is used to determine optimal Earth model parameters (elastic lithosphere thickness, upper and lower mantle viscosities) by calculating the model misfits across the parameter space. The first case study uses the regional RSL database of Vacchi et al. (2018) which spans the Canadian Arctic and East Coast with several hundred sea level index points and limiting points for the early to late Holocene. The second case study uses a global database of several thousand contemporary uplift rates derived from GPS data (Schumacher et al., 2018). For the first case study we find two main features from incorporating LV structures compared to the SS configuration: a decrease in the best scoring misfit and a shift of the misfit distribution in the parameter space to favour a reduced upper mantle viscosity and reduced sensitivity to the lower mantle viscosity.ReferencesAbadi, M., Agarwal, A., Barham, P., et al.: TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems, https://www.tensorflow. org/.Latychev, K., Mitrovica, J. X., Tromp, J., et al.: Glacial isostatic adjustment on 3-D Earth models: a finite-volume formulation, GJI, 161, 421–444, https://doi.org/10.1111/j.1365-246x.2005.02536.x, 2005.Schumacher, M., King, M. A., Rougier, J., et al.: A new global GPS data set for testing and improving modelled GIA uplift rates, GJI, 214, 2164–2176, https://doi.org/10.1093/gji/ggy235, 2018.Vacchi, M., Engelhart, S. E., Nikitina, D., et al.: Postglacial relative sea-level histories along the eastern Canadian coastline, QSR, 201, 124–146, https://doi.org/10.1016/j.quascirev.2018.09.043, 2018.Whitehouse, P. L.: Glacial isostatic adjustment modelling: historical perspectives, recent advances, and future directions, Earth Surface Dynamics, 6, 401–429, https://doi.org/10.5194/esurf-6-401-2018, 2018. |
