Deep Learning Model Automates Oil Shale Organic Analysis

Published in Earth Science Informatics, the study addresses the significant subjectivity and inefficiency inherent in manual maceral identification, a process that has traditionally hindered large-scale, high-precision quantitative analysis in petroleum geology. The framework integrates convolutional neural network-based image classification, semantic segmentation, and Bayesian hyperparameter optimization to enable accurate and reproducible maceral quantification across diverse geological samples.

The study focused on Mesozoic oil shale samples characterized by low thermal maturity and abundant organic microcomponents. A total of 209 oil shale core samples were systematically collected from four distinct basins: Yin’e, Fushun, Songliao, and Tarim. Thin sections prepared from these samples were used to construct a comprehensive semantic segmentation dataset. An additional 72 Yin’e Basin thin sections were reserved for geological evaluation and were strictly excluded from the segmentation training, validation, and test subsets to ensure unbiased model evaluation. This multi-basin approach allows the model to be tested against varied depositional settings and organic matter compositions.

Organic microcomponents, which encompass organic macerals and associated pyrite, were identified under a Zeiss transmitted and reflected fluorescence microscope at 50x magnification. To ensure the accuracy and reliability of the point-counting results, at least 500 points were counted per thin section across multiple fields of view. The dataset was annotated for seven specific classes: Sporinite, Lamalginite, Inertinite, Pyrite, Cutinite, Telalginite, and Vitrinite. Annotations were cross-checked by a second observer with expertise in organic petrology, with any disagreements resolved through joint re-examination under the microscope until consensus was reached.

To mitigate the challenge of insufficient training data, the researchers employed data augmentation techniques including random horizontal flipping, cropping, and blending of white-light and fluorescence images. The dataset was partitioned at the thin-section level rather than the individual image level to prevent data leakage, as adjacent fields of view from a single thin section share petrographic composition and are not statistically independent. This procedure yielded 1,531 images for training, 439 for validation, and 220 for testing, ensuring that the test subset provided a sample-level blind evaluation of model generalization to unseen rock samples.

The semantic segmentation model was trained using a ResNet-50 backbone, which offers robust feature extraction and mitigates the vanishing gradient problem associated with deep networks. Three candidate architectures—FCN, DeepLabV3, and PSPNet—were evaluated under a unified Bayesian optimization protocol. Each architecture completed 50 independent optimization trials using the Tree-structured Parzen Estimator as the surrogate model. FCN-8 S offered the best trade-off between segmentation accuracy and computational efficiency, reaching a pixel accuracy of 90.26% and a mean Intersection over Union of 56.97% on the held-out test set.

Bayesian optimization played a critical role in enhancing model performance by systematically tuning hyperparameters such as learning rate, weight decay, momentum, and the number of training epochs. With default hyperparameters, the model achieved a pixel accuracy of 86.42% and a mean Intersection over Union of 35.58%. After Bayesian optimization, these metrics improved to 92.11% and 58.75%, respectively, on the validation set. This substantial improvement demonstrates the efficacy of the optimization protocol in locating optimal hyperparameter combinations that minimize prediction errors and improve segmentation accuracy.

The per-class results revealed a markedly imbalanced distribution of microcomponents, with vitrinite and lamalginite together accounting for approximately 83.6% of all microcomponent pixels. Misclassified pixels from these rare classes were predominantly assigned to volumetrically dominant classes, consistent with their low recall. This class imbalance highlights a limitation of the current model, as reliable quantification of rare macerals remains challenging despite the high overall pixel accuracy.

To validate the geological consistency of the model output, the segmentation-derived kerogen type index was compared with manually measured indices for the 72 Yin’e Basin thin sections. The model-predicted and manually measured indices exhibited a strong, statistically significant linear correlation with a Pearson r of 0.921 and a mean absolute error of 16.0. The model-derived index also correlated significantly with hydrogen index, total organic carbon, and hydrocarbon generation potential, attaining higher R-squared values than the manually measured indices for these geochemical parameters.

Rock-Eval pyrolysis of the Yin’e Basin samples showed that total organic carbon ranged from 0.53% to 13.6%, averaging 5.26%. Hydrocarbon generation potential ranged from 0.19 to 152.47 mg/g, indicating generally good generation potential. The peak temperature of hydrocarbon cracking ranged from 423 to 442 degrees Celsius, indicating immature to low-maturity organic matter. These geochemical characteristics, combined with the high hydrogen index, suggest that the oil shales are dominated by oil-prone kerogen, highlighting the importance of organic geochemical analysis for evaluating resource potential.

The study also explored the association between the kerogen type index and paleoenvironmental proxies such as Th/K, U/Th, and V/Cr ratios. Type I samples exhibited the lowest Th/K values, indicating fine-grained terrigenous input under weak hydrodynamic conditions that favor organic matter accumulation. Both U/Th and V/Cr ratios increased in more sapropelic samples, consistent with enrichment under reducing conditions and anoxic bottom waters. However, elemental data were available for only 50 samples, so these proxy comparisons are presented as supporting, exploratory evidence rather than primary validation.

Several limitations were acknowledged by the researchers. The dataset is derived from a limited number of wells and thin sections, and the labor-intensive manual annotation limits the scale of the training data. The study focused on low-maturity oil shale samples, and the applicability of the model to higher-maturity samples, where maceral morphologies are altered by thermal maturation, remains to be tested. Furthermore, a dedicated cross-basin validation using samples from a basin entirely excluded from training has not yet been performed.

Despite these limitations, the framework offers a high-throughput approach to kerogen type index estimation, with potential applications to hydrocarbon potential evaluation and paleoenvironmental reconstruction.

The automated identification framework addresses a specific bottleneck in organic petrology: the high diversity and complex structures of organic microcomponents often produce similar morphological features under the microscope. Manual identification is time-consuming, highly subjective, and poorly reproducible, making it difficult to meet the demands of large-scale, high-precision quantitative analysis. While machine learning approaches have been applied to organic-rich shale evaluation, including lithology classification and total organic carbon prediction from geochemical and well-log data, research on automated image segmentation and identification of organic microcomponents in oil shale samples remains limited. In particular, the coupling of automatic identification results with organic geochemical parameters and paleoenvironmental indicators has not been systematically investigated. This study integrates Bayesian optimization into an FCN-based semantic segmentation framework for pixel-level identification and links segmentation-derived results with organic petrological and geochemical analyses.

The semantic segmentation model performs pixel-level inference directly on seven specific classes: Sporinite, Lamalginite, Inertinite, Pyrite, Cutinite, Telalginite, and Vitrinite. The network does not produce intermediate outputs at the maceral-group level. Instead, the maceral groups, comprising vitrinite, exinite, inertinite, and sapropelinite, together with pyrite as an additional non-maceral class, are used solely for the subsequent aggregation of the kerogen type index. This distinction is critical for interpreting the model’s output, as the kerogen type index is a weighted aggregation of maceral-group abundances. The index is calculated using a formula where sapropelinite, exinite, vitrinite, and inertinite percentages are assigned specific weights. The resulting index classifies kerogen types, with values of 80 or higher corresponding to Type I, 40 to less than 80 to Type II1, 0 to less than 40 to Type II2, and values below 0 to Type III.

The sensitivity of the kerogen type index to misclassification depends on the affected abundance and the source and destination groups. For instance, at constant total maceral abundance, reassigning one percentage point of inertinite to vitrinite increases the index by 0.25, whereas reassignment to sapropelinite increases it by 2.0. Misclassification between cutinite and sporinite does not alter the index because both belong to the exinite group. However, a one-percentage-point reassignment of cutinite to sapropelinite increases the index by 0.5, while reassignment to vitrinite decreases it by 1.25. This mechanistic basis explains the coexistence of low rare-class recall and a strong correlation in the aggregate index. The model’s ability to distinguish among microcomponent classes is measured over pixels where both the ground-truth annotation and predicted label belong to one of the seven microcomponent classes, rather than separating organic matter from the mineral matrix.

The geochemical characterization of the Yin’e Basin samples provides context for the model’s geological application. Rock-Eval pyrolysis indicates that the oil shales are characterized by high organic matter abundance, good hydrocarbon generation potential, and immature-to-early-mature thermal maturity. Approximately 84% of samples have total organic carbon greater than 3%, and nearly half exceed 6%. Similarly, approximately 85% of samples have hydrocarbon generation potential greater than 6 mg/g, and 66.4% exceed 20 mg/g. The peak temperature of hydrocarbon cracking ranges from 423 to 442 degrees Celsius, with none exceeding 445 degrees Celsius. This thermal maturity profile, combined with the high hydrogen index, indicates that the organic matter is dominated by oil-prone kerogen. These geochemical parameters serve as the basis for validating the geological relevance of the model-derived kerogen type index.

The study’s framework is designed to offer a high-throughput approach to kerogen type index estimation. By overcoming the subjectivity and inefficiency of manual maceral identification, the automated system enables more consistent and reproducible quantitative analyses. The integration of CNN-based image classification, semantic segmentation, and Bayesian hyperparameter optimization allows for accurate maceral quantification across diverse geological samples. The model’s stable performance on a multi-basin test set suggests its potential for broader application in petroleum geology and organic petrology. Future work will expand the dataset to additional basins and maturity levels, conduct basin-excluded independent validation, and further integrate segmentation-derived quantification with geochemical and paleoenvironmental analyses to refine the model’s utility in resource evaluation.

Subject of Research: Automatic identification and application of organic microcomponents in oil shale based on deep learning models using Bayesian optimization algorithm

Article Title: Automatic identification and application of organic microcomponents in oil shale based on deep learning models using Bayesian optimization algorithm

Article References: Liu, Q., Liu, R., He, T., & Huang, Y. (2026). Automatic identification and application of organic microcomponents in oil shale based on deep learning models using Bayesian optimization algorithm. Earth Science Informatics, 19(12), Article 239. https://doi.org/10.1007/s12145-026-02295-4

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02295-4

Keywords: Automatic, identification, application, organic, microcomponents, shale, based, deep, learning, models, Bayesian, optimization