Prediction of Mineral Content and Petrophysical Parameters in Lacustrine Fine-Grained Mixed Sedimentary Rocks Based on a Physics-Informed Hybrid Deep Learning Framework
Yiming Zhu, Yukun Liu*, Shuya Chen, Yulin Du , Xiang Cheng, Xiaolong Wang
Key Laboratory of Exploration Technologies for Oil and Gas Resources (Yangtze University), Ministry of Education, Wuhan 430100, China.
To address the challenges of severe logging response overlap and quantitative mineral prediction caused by the strong heterogeneity of lacustrine carbonate reservoirs, this study proposes a hybrid framework integrating a physics-informed Gradient Boosting Decision Tree (LightGBM) and a Temporal Convolutional Network (PINN-TCN). The framework employs a sliding-window TCN to capture the deep contextual information of logging sequences, and introduces a truncation strategy along with a physics-constrained loss function in the prediction phase to force the network outputs to comply with the laws of mineral volume closure and density conservation. Coupled with a cosine annealing dynamic weight scheduling strategy, the model achieves a smooth transition from physics-constrained to data-driven optimization, establishing a closed-loop optimization system. Validation using 14,104 data records from the Qianjiang Depression in the Jianghan Oilfield demonstrates that the framework achieves excellent prediction accuracy across three major mineral components, yielding a closure root-mean-square error (RMSE) of only 0.0124, with 99.7% of the predicted values falling within a 5% deviation. Furthermore, by adaptively learning mineral density parameters, the RMSE of solid density prediction is reduced by 73.3%. SHAP (Shapley Additive exPlanations) attribution analysis further confirms that the physical constraints induce a synergistic adjustment of feature contributions within the decision space. This study provides a novel approach for the quantitative evaluation of minerals in complex lithofacies reservoirs, balancing both physical consistency and interpretability.
Download the full text PDF for viewing and using it according to the license of this paper.
Funding
This work was supported by the Natural Science Foundation of Hubei Province, China (Grant No. 2024AFB271).
References
Bagheri, H., Mohebian, R., Moradzadeh, A., & Mehmandost Olya, B. A. (2024). Pore size classification and prediction based on distribution of reservoir fluid volumes utilizing well logs and deep learning algorithm in a complex lithology. Artificial Intelligence in Geosciences, 5, 100094. doi:10.1016/j.aiig.2024.100094
Bodong, Z., Maojin, T., Guohua, Z., & Chenglin, L. (n.d.). An integrating physics-informed neural networks and attention mechanism based intelligent permeability prediction method for shale reservoirs.
Cao, Z., Qin, Z., Luo, S., Chen, F., Huang, K., Dong, X., Meng, L., Su, K., & Wei, K. (2026). Research and insights into the impact of different sampling strategies on machine learning-based lithology identification using well logging data. Artificial Intelligence in Geosciences, 7(2), 100224. doi:10.1016/j.aiig.2026.100224
Guo, J., Wang, C., Wang, Y., Zhang, W., Zhao, Q., Wei, H., & Nie, X. (2026). A shale-oil reservoir multi-mineral inversion method based on well-logging data using an improved random-forest approach: Methodology and application. Applied Computing and Geosciences, 29, 100334. doi:10.1016/j.acags.2026.100334
Haritha, D., Satyavani, N., & Ramesh, A. (2025). Generation of missing well log data with deep learning: CNN-bi-LSTM approach. Journal of Applied Geophysics, 233, 105628. doi:10.1016/j.jappgeo.2025.105628
Hussain, W., Luo, M., Ali, M., Sadiq, I., Kasala, E. E., Aziz, T., & Batool, Z. (2025). Hybrid modeling of deep neural networks and unsupervised machine learning algorithms for missing well log prediction based on geological lithofacies similarities. Journal of Applied Geophysics, 241, 105846. doi:10.1016/j.jappgeo.2025.105846
Kang, Z., Liu, Y., Wang, X., & Du, Y. (2026). A deep learning-based method for well log data reconstruction in marine carbonate reservoirs. Journal of Applied Geophysics, 246, 106106. doi:10.1016/j.jappgeo.2026.106106
Lai, J., Su, Y., Xiao, L., Zhao, F., Bai, T., Li, Y., Li, H., Huang, Y., Wang, G., & Qin, Z. (2024). Application of geophysical well logs in solving geologic issues: Past, present and future prospect. Geoscience Frontiers, 15(3), 101779. doi:10.1016/j.gsf.2024.101779
Li, D., Li, X., Liu, L., He, W., Li, Y., Li, S., Shi, H., & Fan, G. (2025). Prediction on rock strength by mineral composition from machine learning of ECS logs. Energy Geoscience, 6(2), 100386. doi:10.1016/j.engeos.2025.100386
Pang, Q., Chen, C., Li, W., & Pang, S. (2026). Lithology identification from missing well log data: A multi-constraint guided self-supervised diffusion framework. Advanced Engineering Informatics, 74, 104697. doi:10.1016/j.aei.2026.104697
Park, J., Jeong, J., & Holden, E.-J. (2026). Deep-learning-based hybrid model with iterative lithology constraints for the enhanced prediction of missing well-logs. Computers & Geosciences, 209, 106106. doi:10.1016/j.cageo.2026.106106
Pothana, P., & Ling, K. (2025). Physics-integrated neural networks for improved mineral volumes and porosity estimation from geophysical well logs. Energy Geoscience, 6(2), 100410. doi:10.1016/j.engeos.2025.100410
Sun, Y., Pang, S., Li, H., Wang, W., & Qiao, S. (2026). Architectural advancements in dual-contrastive learning: A framework for precision lithology identification from geophysical well logging data. Journal of Applied Geophysics, 251, 106310. doi:10.1016/j.jappgeo.2026.106310
Wu, X., Xu, H., Zhou, H., Wang, L., Jiang, P., & Wu, H. (2024). Improving lithofacies prediction in lacustrine shale by combining deep learning and well log curve morphology in sanzhao sag, songliao basin, China. Computers & Geosciences, 193, 105735. doi:10.1016/j.cageo.2024.105735
Yang, Y.-Q., Qiu, L.-W., Gregg, J., Shi, Z., & Yu, K.-H. (2016). Formation of fine crystalline dolomites in lacustrine carbonates of the eocene sikou depression, bohai bay basin, east China. Petroleum Science, 13(4), 642–656. doi:10.1007/s12182-016-0128-z
SHOW FULL COLUMNS FROM `hl_config_system` [ RunTime:0.001464s ]
SELECT `value` FROM `hl_config_system` WHERE `type` = 'baseweb' AND `stype` = 'sys_view' AND `versions` = 0 LIMIT 1 [ RunTime:0.000565s ]
SELECT * FROM `hl_config_system` WHERE `type` = 'website' [ RunTime:0.000490s ]
SHOW FULL COLUMNS FROM `hl_foot_column` [ RunTime:0.001310s ]
SELECT `id`,`pid`,`title`,`url`,`type`,`link_to`,`link_way` FROM `hl_foot_column` WHERE `versions` = 0 AND `is_show` = 1 ORDER BY `sorted` DESC,`id` ASC [ RunTime:0.000492s ]
SELECT `value`,`stype` FROM `hl_config_system` WHERE `type` = 'baseweb' AND `versions` = 0 [ RunTime:0.000399s ]
SHOW FULL COLUMNS FROM `hl_journal_subject` [ RunTime:0.001024s ]
SELECT `id`,`subject_name` FROM `hl_journal_subject` WHERE `versions` = 0 ORDER BY `sort` DESC,`id` DESC [ RunTime:0.000300s ]
SHOW FULL COLUMNS FROM `hl_journal` [ RunTime:0.001748s ]
SELECT `jid`,`name` FROM `hl_journal` WHERE `versions` = 0 AND `isdelete` = 0 ORDER BY `is_recommend` DESC,`sort` DESC,`jid` DESC [ RunTime:0.000903s ]
SHOW FULL COLUMNS FROM `hl_paper` [ RunTime:0.002498s ]
SELECT * FROM `hl_paper` WHERE `id` = 3556 LIMIT 1 [ RunTime:0.000555s ]
UPDATE `hl_paper` SET `cnum` = `cnum` + 1 WHERE `id` = 3556 [ RunTime:0.003182s ]
SELECT `name` FROM `hl_journal` WHERE `jid` = 65 LIMIT 1 [ RunTime:0.000773s ]
SHOW FULL COLUMNS FROM `hl_journal_period` [ RunTime:0.001813s ]
SELECT `pid`,`name`,`cyear`,`imgurl` FROM `hl_journal_period` WHERE `isdelete` = 0 AND `id` = 477 LIMIT 1 [ RunTime:0.000520s ]
SELECT `pid`,`name`,`cyear`,`imgurl` FROM `hl_journal_period` WHERE `id` = 397 LIMIT 1 [ RunTime:0.000379s ]
SHOW FULL COLUMNS FROM `hl_paper_tag` [ RunTime:0.001845s ]
SELECT * FROM `hl_paper_tag` WHERE `id` IN (247,337) ORDER BY `sort` DESC,`id` ASC [ RunTime:0.000428s ]
SELECT * FROM `hl_journal` WHERE `jid` = 65 LIMIT 1 [ RunTime:0.000762s ]
SELECT `pid`,`name`,`type`,`cyear`,`id` FROM `hl_journal_period` WHERE `jid` = 65 AND `isdelete` = 0 ORDER BY `cyear` DESC,`type` ASC,`name` ASC,`ctime` ASC [ RunTime:0.000816s ]
SHOW FULL COLUMNS FROM `hl_journal_menu` [ RunTime:0.001624s ]
SELECT `id`,`column_title`,`column_content` FROM `hl_journal_menu` WHERE `jid` = 65 [ RunTime:0.001871s ]
SHOW FULL COLUMNS FROM `hl_paper_menu` [ RunTime:0.001607s ]
SELECT `column_title`,`column_content` FROM `hl_paper_menu` WHERE `pid` = 3556 [ RunTime:0.003899s ]
UPDATE `hl_journal` SET `sections_number` = `sections_number` + 1 WHERE `jid` = 65 [ RunTime:0.002420s ]
