Text-Guided Subsurface Model Generation for Geological Realism
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Generative models struggle to generate geologically realistic subsurface models due to the unique distribution and semantic understanding required for seismic imagery, which differs significantly from general-purpose text-to-image models.
Innovation Solution
A system that integrates text-to-image machine learning technologies to generate geologically realistic subsurface models by training on text-model pairs, using a machine learning model to encode text prompts into embeddings and produce accurate subsurface models.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If general-purpose text-to-image models are used for subsurface model generation, then the models can generate visual content from text descriptions, but the generated models lack geological realism and semantic accuracy
Solution Approach 1:
The patent changes the training parameters and data distribution by training the generative model specifically on seismic imagery and geological data rather than general images. This involves adjusting the model's learning parameters to match the unique statistical properties, semantic structures, and visual characteristics of subsurface geological data, enabling the model to generate geologically realistic models while maintaining text-to-image generation capability
Solution Approach 2:
The patent segments the general text-to-image generation task into domain-specific components by creating a specialized model architecture and training pipeline tailored for geological data. This involves separating the general generative capabilities from the domain-specific knowledge, allowing the model to adapt to the unique requirements of subsurface model generation while preserving the core text-to-image functionality
2Ease of operation
If standard text-to-image models are trained on general imagery, then the models can understand and combine objects semantically, but they cannot generalize to the subsurface domain with its unique data distribution
Solution Approach 1:
The patent creates a universal text-to-image generation framework that can function across multiple domains. By training the model on domain-specific seismic imagery while maintaining the general text-to-image architecture, the model achieves multi-functionality: it retains the ability to understand and combine objects semantically from general training, while simultaneously adapting to the unique data distribution and semantic structures of subsurface geological data
3Productivity
If general-purpose models are used without domain-specific training, then the models can process text descriptions efficiently, but they fail to capture the stochastic and probabilistic nature of geological data
Solution Approach 1:
The patent applies preliminary action by pre-training the generative model on extensive seismic imagery datasets before deploying it for specific geological tasks. This preliminary training phase allows the model to learn the stochastic and probabilistic patterns inherent in geological data, including the natural variability, noise characteristics, and statistical distributions of subsurface features, thereby improving reliability while maintaining processing efficiency
Data Source
AI summary
A method, system, and computer program product are provided for generating a subsurface model by an image generating machine learning model (MLM). A text prompt is received describing a subsurface geological feature. A text encoder encodes the text prompt into a text embedding. The text embedding is processed by the image generating MLM, specifically trained on pairs of text descriptions and subsurface models, to generate subsurface models from text. The image generating MLM outputs a subsurface model comprising the subsurface geological feature that corresponds to the text prompt.