Advancement in Seismic Data Conditioning Using Tailored Deep Neural Networks

Abstract This paper presents a custom Machine Learning approach for reducing seismic noise while improving resolution and features identification. The seismic data enhancement created via this approach has positive impact on every stage of the oil-field life cycle, however, it is particularly advantageous during the exploration phase, where high quality seismic data is often unavailable from traditional solutions due to several factors. The high-quality seismic data output from the custom Machine Learning approach enables more accurate identification of hydrocarbon reservoir potentials, reduces prospect risk, provides basis for optimized well placements, and would ultimately contribute to higher exploration and exploitation success rates. Overall, this solution approach involves autogenerating Deep Neural Network models that are individually tailored to remove noise from input seismic data, while improving resolution and preserving edges. The process consists of five stages: pre-processing, features extraction, neural network model training, post-processing, and results evaluation. Training data is generated by perturbing the input seismic data to create a super-noisy version of the data. The deep neural network then creates a model that removes noise from the perturbed data to recover the original input data. Thereafter, the same noise removal model is applied to the original unperturbed seismic data to generate a noise-free version of the original seismic data. Since the training data is generated from the original seismic itself, the noise removing model is tailored for each seismic data volume. The process does not require well data input. Output seismic volumes from the Machine Learning Conditioning process in this project had better signal to noise ratio compared to the input. Seismic events were realistically smoothened and sharpened without smearing edges, thus improving horizon and fault interpretation. Some hitherto hidden reflector relations were revealed in the output seismic data, thus improving stratigraphic interpretation. Seismic attributes also had better quality from the output seismic data. The process is automated, so it saves time and makes it possible to efficiently work with multiple datasets to reduce interpretation uncertainty. As the industry moves towards exploring and developing more challenging green reservoirs and redeveloping brown reservoirs, having the cost-effective means provided by this solution helps in generating more detailed insight from existing seismic data volumes, thereby creating multiplied value. The new fully automated method developed for conditioning seismic data in this project is unlike other solutions in that it learns noise patterns from the input data and removes them in the same run, making it more versatile than pretrained models that only work in specific settings. As the process does not require well data input, it is useful right from the earliest phases of exploration. Results are more accurate than traditional data conditioning methods, and the process is extremely quick compared to the alternative of having to reprocess the seismic data.