Local Rock Property Estimation Using Transfer Learning for Drilling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing petrophysical evaluations face challenges in accurately estimating rock properties due to the complexity and heterogeneity of geological formations, necessitating geologically-specific models that are often time-consuming and prone to instability with limited data.
Innovation Solution
A method involving transfer learning using a pre-trained global model to generate local weights for a local model, optimizing drilling parameters and petrophysical property estimation by fine-tuning or adding new layers with local data, reducing the need for extensive retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a local model is trained from scratch to improve accuracy for specific geological formations, then estimation accuracy is improved, but training time and computational resources increase significantly
Solution Approach 1:
A global model is pre-trained on a large, diverse dataset encompassing multiple geological formations before being adapted to local conditions. This preliminary training establishes a robust foundation that can be efficiently fine-tuned for specific formations, avoiding the need to train from scratch each time.
Solution Approach 2:
The system transitions from a single global model to multiple formation-specific local models, each tailored to the unique characteristics of specific geological formations. This allows each local model to achieve high accuracy for its target formation while leveraging the general knowledge from the global model through transfer learning.
2Reliability
If formation-specific local models are developed to improve estimation accuracy, then model reliability is improved, but data availability and model stability deteriorate due to limited local data
Solution Approach 1:
The global model serves as an intermediary that bridges the gap between limited local data and the need for reliable local models. It transfers knowledge from abundant global data to supplement the scarce local data, enabling robust local model training even when local datasets are small.
Solution Approach 2:
The system adjusts model parameters dynamically by fine-tuning the pre-trained global model using local data. This allows the model to adapt its weights and biases to local conditions while maintaining the stability provided by the extensively trained global foundation, achieving reliability with limited data.
3Ease of operation
If conventional global models are used for all formations to simplify the process, then ease of operation is improved, but measurement precision deteriorates due to geological heterogeneity
Solution Approach 1:
The system segments the single global model into multiple formation-specific local models, each optimized for particular geological formations. This segmentation allows each model to capture formation-specific characteristics and relationships, improving accuracy while maintaining a standardized approach through the transfer learning framework.
Solution Approach 2:
The transfer learning framework provides universality by enabling a single pre-trained global model to be efficiently adapted to multiple different geological formations. The same methodology and architecture can be applied across various formations, maintaining ease of operation while achieving formation-specific precision.
Data Source
AI summary
Systems and methods for performing local optimization of rock property estimation in geological formations are provided. A method includes: drilling into a rock formation using first drilling parameters, acquiring local data from a first sample from the drilling, acquiring test data from a second sample, selecting a local model input and output, receiving a pre-trained global model including a global model input and output, accessing the global model to extract global weights for global neuron layers, passing the global weights to a local model, training the local model with the local data using the passed global weights to generate local weight(s) corresponding to local neuron layer(s), feeding the test data into the trained local model to generate a prediction output, and based on the prediction: generating second drilling parameters to optimize drilling of the rock formation, and drilling into the rock formation using the second drilling parameters.