× close
A petrophysics-based residual neural network for multitask reservoir parameter prediction using data mechanism-driven loss functions.Credit: Rongbo Shao et al.
Data-driven artificial intelligence, such as deep learning and reinforcement learning, has powerful data analysis capabilities. These techniques enable statistical and probabilistic analysis of data and facilitate the mapping of relationships between inputs and outputs without relying on predetermined physical assumptions.
Central to the training process for data-driven models is the utilization of a loss function that calculates the difference between the model’s output and the desired target result (label). The optimizer then adjusts the model’s parameters based on the loss function to minimize the difference between the output and the label.
Geophysical logging, on the other hand, involves a wealth of disciplinary knowledge, mathematical models, and physical models. Relying solely on data-driven models may produce results that are inconsistent with established knowledge. Additionally, training data with uneven distribution or subjective labels can also affect the performance of data-driven models.
Recent research published in Artificial intelligence in earth sciences reported the implementation of constraints for training a data-driven machine learning model using logging response functions in a logging reservoir parameter prediction task.
“Our model, called Petrophysics Informed Neural Network (PINN), integrates petrophysics constraints into the loss function to guide training,” said Dr. Longbo Hsiao, lead author of the study. states. Candidate at China University of Petroleum, Beijing. “During model training, if the model output differs from the rock physics knowledge, the loss function is penalized by the rock physics constraints. This causes the output to be closer to the theoretical value and the labeling for training the model It reduces the impact of errors.”
Additionally, this approach helps identify the correct relationships from the training data, especially when the sample size is small.
“To make the influence of the mechanistic model more flexible in the machine learning model, we introduced tolerances and petrophysical constraint weights,” Shao elaborates. “We evaluated the ability of his PINN model to predict reservoir parameters using measured data.”
Shao et al. found that this model has improved accuracy and robustness compared to purely data-driven models. Nevertheless, the researchers noted that the selection of petrophysical constraint weights and tolerances remains subjective and therefore requires further investigation.
“The integration of data-driven AI models and knowledge-driven mechanism models is a promising research area. The success of the PINN model in well logging is a great step forward,” said corresponding author Professor Lizhi Xiao from China University of Petroleum. “Yes,” he said, emphasizing the importance of this research. For earth science in this direction. ”
Emphasizing the need for continuous improvement, Xiao said, “Continued challenges exist in the selection of petrophysical constraint weights and tolerances, and the adaptability of domain knowledge to different geological formations. .Furthermore, the quality of the dataset is critical for AI applications.”Geophysical Logging. High quality, quantitative, comprehensive, and publicly available well logging datasets are required. ”
For more information:
Rongbo Shao et al., Reservoir Assessment Using Petrophysics-Based Machine Learning: A Case Study, Artificial intelligence in earth sciences (2024). DOI: 10.1016/j.aiig.2024.100070
Provided by: KeAi Communications Co., Ltd.
