Towards the prediction of elastic (seismic) anisotropy in the upper mantle using supervised deep-learning

TOMMASI, ANDREA; CERPA, NESTOR; CARAZO, FERNANDO DIEGO; SIGNORELLI, JAVIER

Viena

Congreso; EGU General Assembly; 2024

Austrian Geoscience Society

Both elastic and viscoplastic behaviors of the Earth’s upper mantle are highly anisotropic, because olivine, the orthorhombic phase composing 60-80% of the mantle, has a strong intrinsic viscoplastic anisotropy and develops strong textures. Predicting the evolution of olivine anisotropy with strain is essential to: (1) probe indirectly the deformation in the convecting mantle with seismic measurements and (2) accounting for the deformation history when simulating the long-term dynamics of the Earth. However, traditional micro-mechanical approaches to model the evolution of this texture-induced anisotropies are too memory-costly and time consuming for coupling into geodynamical simulations. To speed up the prediction of elastic anisotropy in the mantle, we developed deep-learning (DL) surrogates trained on a synthetic database built with viscoplastic self-consistent simulations of texture evolution of olivine polycrystals in typical 2D geodynamical flows. A first challenge was the choice of memory-saving representations of the texture. Training the DL models on the evolution of the elastic tensor components avoids the need of saving the texture itself. Yet, the major challenge has been to prevent error compounding in a recursive-prediction scheme – where a model prediction at a given time step becomes the input for the next one - to evaluate the anisotropy evolution along a flow line. We implemented multilayer feed forward (FFNN), ensemble and transformer neural networks, obtaining the best efficiency/accuracy ratio for the FFNN. The results highlight the importance of (1) the standardization of the outputs in the training stage to avoid overfitting in predictions, (2) the statistical characteristics of the strain histories in the training database, and (3) the influence of non monotonic strain histories on error propagation. Predictions for complex unseen strain histories are accurate, much more time-efficient and memory-costly than the traditional micro-mechanical models. Our work opens thus new avenues for modeling the strain-controlled evolution of mechanical anisotropy in the Earth’s mantle.