Pretrained Foundation Model Enables Rapid Seismic Interpretation

_ This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper SPE 227546, “PRISM: A Pretrained Foundation Model for Rapid Seismic Interpretation,” by Haibin Di and Arvind Sharma, SLB. The paper has not been peer-reviewed. _ While robust seismic interpretation and modeling is essential in successful subsurface mapping and characterization, it remains challenging because of its strong dependency on data quality, domain knowledge, and expert supervision. Pretraining of a large seismic foundation model (FM) can provide rich and reliable representation of diverse seismic patterns observed on post-stack seismic images and adopt it into multiple downstream workflows. This paper presents a self-supervised approach for training such a seismic FM and demonstrates its applications. Seismic FM Model Architecture. The proposed seismic FM is made of an encoder, decoder, three refiners, and five taskers, which enable multitasking training from a shared embedding space. Specifically, the encoder generates a rich embedding from input seismic and of a vision-transformer-large architecture from the Dino-v2 model pretrained on natural images as the starting checkpoint. The decoder upscales the encoder embedding to a set of features with the same dimension as the input seismic and is made of four upsampling layers trained from scratch. The three refiners compress the decoder features into a smaller dimension from three perspectives: amplitude, waveform, and interpolation. Each is made of four convolutional layers and is trained from scratch. The five taskers are for five different tasks: denoising, energy, phase, and lateral and vertical interpolation. Each tasker is made of two convolutional layers and is trained from scratch. Training Data. The data for training the proposed seismic FM are 2D tiles retrieved from both publicly available real seismic and synthetic seismic. Specifically, the former covers the areas of North Sea, Netherlands, New Zealand, and Australia. The synthetic seismic is simulated by first generating realistic structural models, then converting them into velocity models and convolving the derived reflectivity with a wavelet, plus post-processing for such issues as random noise contaminations. For training the model in a self-supervised manner, each tile is further processed by adding noises, masking out vertical and horizontal randomly, and applying complex seismic trace analysis to derive instantaneous energy and phase. Adoption Scenarios. Fig. 2 of the complete paper shows three scenarios for adopting the trained seismic FM, with selected use causes. These include the following: - Direct inference either runs the tasker heads existing in the seismic FM or applies operators to the latent space (encoder embedding or decoder features). These use cases include 2D to 3D interpolation, 4D anomaly detection, and geofeature extraction. - Instantaneous learning extracts the latent space from the seismic FM, feeds them with additional information if available into simple machine-learning schemes such as random forest and multilayer perceptron (MLP), and builds corresponding simple machine-learning models using the provided labels instantaneously. Such learning allows for integrating seismic and wells for marker propagation and property inversion efficiently. - Light fine-tuning attaches a small downstream convolutional neural network (CNN) to the seismic FM blocks such as the encoder or decoder, feeds it with the seismic FM latent space together with additional information if available, and trains only the attached small CNN using the provided labels within a short training period (minutes). One use case is one- or few-shot seismic-image segmentation.