Ensemble of Neural Networks Utilizing Seismic Attributes for Rock‐Property Inversion With Uncertainty Estimation.

Seismic inversion holds significant importance across various domains of geoscience and engineering, including the characterization of energy resource reservoirs, the assessment of polluted sites, and CO2 storage. It is a process of estimating rock properties from seismic data that is inherently uncertain, nonlinear, non‐unique, and highly challenging. Using multiple seismic attributes increases the size of the data, requiring considerable processing resources and time. However, deep learning can accurately fit quantities of nonlinear variables, making it an excellent method for predicting spatially distributed subsurface properties. We trained some multi‐output regression neural networks to carry out porosity inversion from seismic data. We initially computed a series of seismic attributes and generated the corresponding porosity using interpreted horizons, well logs, and seismic data. Subsequently, we proposed a technique to identify the most relevant seismic attributes for porosity inversion. Because our networks work as stochastic modeling entities, we created a weight‐averaging ensemble approach to build a strong model with the highest level of accuracy. We combined realizations from baseline entities, considering their respective performance levels. Using the statistics between these realizations and the robust model, we determined the degree of uncertainty associated with the outcome. We found an R2 of 0.993 and an MAE of 0.00112 in the F3 block offshore the Netherlands, proving the method's effectiveness. The mean porosity was 0.175193, compared to 0.175626 from a reference model, and the mean uncertainty was ±0.0008998. Plain Language Summary: Understanding subsurface properties, such as porosity, is crucial for various applications, such as identifying energy resources and determining suitable locations for CO2 storage. Typically, obtaining the latter involves using the inversion of seismic records, which can be a challenging process with multiple possible outcomes. Using many seismic measures leads to data size expansion and challenges the processing, necessitating additional time and resources. Deep learning is highly effective at accurately measuring subsurface features when analyzing large amounts of random data. As a result, we performed calculations on seismic measurements and used horizons and borehole records to collect porosity samples. We have put forward a technique for identifying the most pertinent metrics. Afterward, we conducted training on multiple networks to make predictions about porosity. We developed several models and created a "weight‐averaging ensemble" to account for the random nature of the networks. This considers the power of baseline models when incorporating realizations, resulting in a strong, more accurate model. We used statistical analysis to determine the level of uncertainty in the approach. The proposed method for measuring porosity in the F3‐Netherlands offshore block is faster than traditional approaches and produces results similar to a reference model. Key Points: This study developed an approach to identify pertinent seismic attributes for porosity inversion with uncertainty estimatesThese attributes are used to train an array of distinct networks used to generate multiple realizationsA grid search is developed to find the best weights assigned to these realizations, which are combined to achieve the most precise result [ABSTRACT FROM AUTHOR]