Deep Learning in Salt Interpretation from R&D to Deployment: Challenges and Lessons Learned Pandu Devarakota 1 , Apurva Gala 2 , Zhenggang Li 3 , Engin Alkan 2 , Yihua Cai 2 , John Kimbro 2 , Dean Knott 1 , Jeff Moore 1 , Gislain Madiba 2 1 Shell Global Solutions US Inc, 2 Shell International Exploration and Production Company, 3 University of Houston Summary With the evolution of the deep learning eco-system and the availability of open-source software packages such as TensorFlow and PyTorch, creating a quick proof-of-concept (PoC) has become a straightforward task. However, based on our experience, we contend that transitioning a project from PoC to Deployment is a difficult process in which the team must systematically consider a plethora of design and data- centric choices, which we refer to as R&D challenges. Some of the R&D challenges toward developing a successful deep learning-based system in the seismic processing domain are presented in this brief abstract. Recommendations on how to mitigate these challenges are discussed on a real-world example of automating salt interpretation. Introduction Salt interpretation plays a critical role in velocity model building in both exploration and development fields. It is a time-consuming effort that requires key domain expertise, and it is critical for imaging complex salt formations in subsalt and pre-salt rich prospects such as the Gulf of Mexico, Brazil, and the North Sea prospects. Typical imaging workflow is an iterative process in which the initial sediment only velocity model is iteratively refined using labor intensive salt picking and compute intensive migration steps. Recently, many authors have used deep learning (DL) to accelerate the salt interpretation step demonstrating that DL can play a significant role in model building step and can substantially reduce the overall seismic imaging workflow cycle time (Shi 2019, Zeng 2019, Consolvo 2020, Sen 2020, Kaul 2021). Even though these findings are encouraging and have the potential to truly transform the way both processing and imaging workflows are performed, deep learning has yet to be widely adopted in production settings in the seismic processing domain. There are numerous research and development challenges that must be addressed in order to produce a generalizable and acceptable solution. In this abstract, we discuss some of those R&D challenges towards developing a successful deep learning-based system to automatically interpret salt geometry, build and render velocity model to run migration with minimal human intervention. As an example, we focus on detecting top of salt (TOS) which is less complex compared to other parts such as base salt overhang (BSOH), top salt overhang (TSOH) and finally base of salt (BOS). We first emphasize the importance of developing systematically high-quality large datasets and its impact on addressing generalization issues. We then discuss why brute force neural network search would fail and how focusing on false positives and false negatives can help in the development of unique approaches. In the second part of the abstract, we shift the focus on addressing challenges that arise during the post- deployment phase. As it is generally known, once these DL models are deployed, they gradually encounter concept drift. Concept drift refers to the deterioration of model’s inference quality as the distribution of data on which the model was trained begins to deviate from the data in the real-world settings. High data variability between training and testing is expected as the models are often trained on legacy datasets and applied to data from new terrains which are acquired and processed through the most recent acquisition and processing techniques. Simple retraining of models on new and old data is both time consuming and it will also lead to the so-called catastrophic forgetting, in which the trained network loses the previously learnt knowledge. To address this problem, we investigated a novel method based on the sharing partial network approach (Sarwar 2020). The key idea of partial network sharing is the unique ‘branch-and-merge’ technique, which allows the network to learn unique knowledge from new data by training new branch without any performance loss in old branch (trained on original training data). The proposed incremental learning approach is tested on a top-salt interpretation task on field datasets from Brazil play. We demonstrate that the incremental training improves time/compute efficiency while preventing the risk of concept drift (Gala 2022). 10.1190/image2022-3751725.1 Page 2989 Second International Meeting for Applied Geoscience & Energy © 2022 Society of Exploration Geophysicists and the American Association of Petroleum Geologists Downloaded 08/17/22 to 35.175.214.152. Redistribution subject to SEG license or copyright; see Terms of Use at http://library.seg.org/page/policies/terms DOI:10.1190/image2022-3751725.1