PythonMachine LearningPyTorch2023

Digitizer: A Synthetic Dataset for Well-Log Analysis

Event Year

2023

BibTeX / Citation

@inbook{Nasim_2024, title={Digitizer: A Synthetic Dataset for Well-Log Analysis}, ISBN={9783031510236}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/978-3-031-51023-6_9}, DOI={10.1007/978-3-031-51023-6_9}, booktitle={Image Analysis and Processing - ICIAP 2023 Workshops}, publisher={Springer Nature Switzerland}, author={Nasim, M. Quamer and Patwardhan, Narendra and Ali, Javed and Maiti, Tannistha and Marrone, Stefano and Singh, Tarry and Sansone, Carlo}, year={2024}, pages={97–105} }

Visit Official Page

Manuscript Preview

Open Fullscreen

Executive Summary

Raster well-log images are digital representations of paper copies that retain the original analog data gathered during subsurface drilling. Geologists heavily rely on these images to interpret well-log curves and gain insights into the geological formations beneath the surface. However, manually extracting and analyzing data from these images is time-consuming and demanding. To tackle these challenges, researchers increasingly turn to computer vision and machine learning techniques to assist in the analysis process. Nonetheless, developing such approaches, mainly those dependent on machine learning requires a sufficient number of accurately labelled samples for model training and fine-tuning. Unfortunately, this is not a straightforward task, as existing datasets are derived from scanned hand-compiled paper copies, resulting in digital images that suffer from noise and errors. Furthermore, these samples only represent images and not the digital signals of the measured natural phenomena. To overcome these obstacles, we present a new synthetic dataset that includes both images and digital signals of well-logs. This dataset aims to facilitate more effective and accurate analysis techniques, addressing the limitations of current methods. By utilizing this dataset, researchers and practitioners can develop solutions that mitigate the shortcomings of existing methods, ultimately leading to more reliable and precise results in interpreting well-log curves and understanding subsurface geological formations.

Related Works

Digitization of Raster Logs: A Deep Learning Application

Raster well logs are digital representations of well logs data and have been generated over the years. Raster digital well logs represent bitmaps of the log image in a rectangular array of black and white dots called pixels. The white pixels are represented by zeros, and the black pixels are represented by ones. Well log data saved as depth-calibrated raster images provide an economic alternative to digital formats for preserving this valuable information into the future (Cisco, 1996). Although often discarded after vectorization, raster imaged well logs may be the key to a global computer-readable format for legacy hardcopy data. This legacy data is stored on multiple media and contain information for a variety of applications in addition to resource exploration and development, such as environmental protection, water management, global change studies, and basic and applied research. Experts revisit and study the raster mostly manually or with software applications that still require tremendous amount of manual input. Besides the loss of thousands of person-hours, this process is both erroneous and tedious. To digitize these raster logs and efficiently use them in conventional as well as unconventional analysis one needs to buy a costly digitizer that are not only a manual and time-consuming task but also there is a hidden technical debt since enterprises stand to lose more money in additional servicing and consulting charges. We propose a novel SOTA (state-of-the-art) AI model VeerNet, powered by our novel and cutting- edge deep learning technologies to address these challenges. We use an attention mechanism, a novel technique that mimics human-like cognitive attention. The crux is that the network devotes more focus to that small but most important part of the data in an encoder-decoder architecture. This architecture efficiently classifies the grid and the curve with an accuracy > 94% for dice loss.

Joint Optimization of Lithology and Petrophysical Parameters in Athabasca Oil Sands Using Self-Attention Mechanism

Precisely determining lithology and petrophysical parameters, including core-calibrated porosity and water saturation, is crucial for reservoir characterization. The traditional method of manually interpreting well-log data is not only time-consuming but also prone to human errors. To address these challenges in identifying lithology and estimating petrophysical parameters in Athabasca Oil Sands region, this study introduces a novel solution using the AutoRegressive Vision Transformer (ARViT) model for accurate prediction. ARViT improves upon the ViT framework by integrating sequential dependencies into its output. The self-attention mechanism and auto-regression are key features that enable ARViT to systematically process data, capturing detailed spatial dependencies in well-log data. This empowers the model to discern subtle spatial and temporal relationships among various geophysical measurements. In essence, ARViT's incorporation of sequential information through its auto-regression mechanism on top of ViT enhances its ability to comprehensively model complex relationships within well-log data. In this study, a multitask learning approach is embraced to enhance the model's interpretability and efficiency. This methodology involves optimizing the model's performance across multiple tasks simultaneously. By doing so, the model gains a broader understanding of diverse tasks and benefits from shared knowledge and features across these tasks. This collaborative optimization contributes to a more robust and versatile model, ultimately improving its overall perflormance and interpretability. To evaluate the effectiveness of the ARViT model, we conducted a comprehensive series of experiments and comparative analyses, contrasting its performance with conventional artificial neural networks (ANN), Long Short-Term Memory (LSTM), and ViT models. Furthermore, to illustrate the versatility of ARViT, we apply Low-Rank Adaptation (LoRA) to a different, smaller dataset of well-log, showcasing its ability to adapt effectively to various geological contexts. LoRA becomes particularly crucial in this context, as it not only enhances the model's adaptability but also plays a vital role in reducing the number of trainable parameters. This reduction not only contributes to computational efficiency but is essential for preventing overfitting and ensuring optimal performance across different datasets. Our findings demonstrate the consistent superiority of ARViT over ANN, LSTM, and ViT in accurately estimating lithological and petrophysical parameters. This is highlighted by ARViT's remarkable Lithological Accuracy of 96.51%, surpassing the baseline ANN's 73.18%, LSTM's 89.80%, and ViT's 93.23%. The substantial reduction in Mean Squared Error (MSE) for porosity, decreasing from 0.0007 (ANN) to 0.0004 (ARViT), and water saturation, decreasing from 0.022 (ANN) to 0.005 (ARViT), further emphasizes ARViT's exceptional performance in providing precise and reliable predictions across various metrics. The application of LoRA yields notable enhancements in ARViT's performance metrics. Specifically, in terms of Lithology Accuracy, ARViT-LoRA showcases a significant improvement, soaring from 88.74% (ARViT-Scratch) to an impressive 97.22%. Additionally, the implementation of LoRA resulted in a significant reduction of GPU consumption by 25%. While lithology prediction has been a well-explored field, ARViT distinguishes itself through its exclusive combination of features, encompassing a self-attention mechanism, auto-regressive nature, and multitask approach, coupled with effective fine-tuning using LoRA. This unique combination positions ARViT as a valuable tool for addressing intricate challenges of lithology prediction and petrophysical parameter estimation.