Estimating porosity distribution is essential for hydrocarbon exploration, carbon capture and storage, and geothermal energy. Traditional workflows, which rely on seismic inversion and rock physics equations, are time-intensive, parameter-dependent, and sensitive to variations in mineralogy and fluid composition. On the other hand, workflows based on machine learning algorithms in the state of the art are also limited by their reliance on specific rock physics equations that may oversimplify geologic complexity. In addition, there are assumptions of data availability in the time domain and requirements for multifrequency seismic inputs, which are not commonly accessible. This study proposes a workflow that uses 1D convolutional neural networks to directly estimate porosity from poststack seismic data, addressing the aforementioned traditional limitations. The workflow implements a custom loss function during training that incorporates porosity derived from well logs. Our proposed workflow outperformed alternative machine learning approaches for porosity estimation from seismic data in both the time and depth domains, using synthetic and field data. Experiments on synthetic data demonstrated its superior predictive capability compared to conventional and state-of-the-art methods. Moreover, evaluations on field data achieved a determination coefficient (R 2 ) of 0.860 on the test set, an outcome reinforced by five-fold cross-validation. This performance shows the network's robust generalization, resilience to data variability, and stability under moderate noise, yielding interpretable results. The network was trained efficiently in less time by requiring fewer trainable parameters, significantly reducing the time required compared to conventional and state-of-the-art workflows.
Cordoba-Castillo et al. (Mon,) studied this question.