Скачать с ютуб MLDADS 2021 - Real time probabilistic inversion of DNN based DeepEM model accounting for model error в хорошем качестве

MLDADS 2021 - Real time probabilistic inversion of DNN based DeepEM model accounting for model error

Presentation by Muzammil Hussain Rammay for the Data Learning working group on ‘Real-time probabilistic inversion of DNN-based DeepEM model while accounting for model error’. This presentation was recorded for MLDADS 2021 - ICCS 2021. Authors included in this work are: Muzammil Hussain Rammay, Sergey Alyaev, Ahmed H. Elsheikh and Reidar Brumer Bratvold. MLDADS 2021: https://www.imperial.ac.uk/events/124... Data Learning Working Group: https://sites.google.com/view/rossell... Abstract: Deep Neural Networks (DNNs) are becoming go-to methods for fast approximation of complex systems which have been traditionally modelled by PDE solvers. For instance, DNNs showed relatively good approximation of the Maxwell's equations required for modeling deep electromagnetic (DeepEM) logging while drilling measurements [Alyaev et al., 2020]. Fast approximations are specifically important to the fields where real-time inversion is required. In drilling operations, the real-time interpretation of subsurface measurements bundled with estimation of relevant subsurface uncertainties could add a significant value by intentionally correcting the well path in real time (known as geosteering). While one can also approximate the inverse operator directly by Deep Learning [Shahriari et al., 2020], recovering relevant uncertainties is non-trivial and in many cases these approximations come short because the inverse problem is ill-posed. Bayesian algorithms could be useful for real-time inversion because of the flexibility to account for non-uniqueness and uncertainties. Among those Bayesian algorithms, iterative ensemble smoothers could be the best choice for real time inversion due to the relatively low computational cost and the parallel nature of the algorithm [Chen and Oliver, 2012]. While significant efforts are usually made to ensure the accuracy of the Deep Learning models, it is widely known that the DNNs contain some type of model-errors in the regions not covered by the training data, which are unknown and training specific. When the Deep Learning models are inverted, the effects of the model errors could be smeared by adjusting the input parameters to match the observations. This results in a bias estimated input parameters and as a consequence might result in a bad quality geosteering operation. In this communication, we evaluate the performance of the probabilistic real-time inversion of a Deep Learning model on the example of inverting the DeepEM geosteering measurements using iterative ensemble smoothers. During the inversion we estimate the boundary positions and resistivities of a layer-cake geological model as well as their associated uncertainties. Such joint inversion of geometry and properties is known to produce an ill-posed inverse problem with local minima. In particular, we focus on model error as one of the main challenges associated with the inversion of Deep Learning models. For this purpose we evaluate two different types of iterative ensemble smoother: the Classical and Flexible ES-MDA [Rammay et al., 2020]. ES-MDA can take into account highly non-linear nature of the Deep learning model and measurement errors, however it does not account for model errors. Our implementation of Flexible ES-MDA, on the other hand, takes into account model error during the probabilistic inversion of DNN model by analysing the change in the residuals during the iterative inversion. We observe that the Flexible ES-MDA has the capability to reduce the effect of model-bias by capturing the unknown model-errors, thus improving the quality of the estimated input parameters for Geosteering. Moreover, we describe the framework for identification of the multi-modality of the real time inversion of Deep learning models using vanilla inversion and possible solutions to alleviate it in real time. The proposed methodology provides a real-time probabilistic inversion framework for Deep Learing models, which accounts for model errors and non-uniqueness. We observe that both iterative ensemble smoothers provide exact estimates when the problem is well posed and has no model errors. When the model errors are present however, the Flexible ES-MDA avoids erroneous convergence to a wrong solution, and preserves a wider posterior which covers the true solution and the true data. Furthermore, in the extreme cases we can detect problems dominated by local minima by comparing the inversion results between the Classical and the Flexible ES-MDA. These issues can be avoided by using informed prior or restarting the inversion with different priors.