Open Access

ARTICLE

Formation Pore Pressure Detection Using a Hybrid Model of Convolutional Neural Network and Machine Learning under Physical Constraints

Xinniu Xu¹, Jinde Li², Hong Huang¹, Guangfu Cao¹, Yuhe Shi², Biao Ruan¹, Daojin Ge¹, Hu Yang^2,*
1 Exploration Division, Xinjiang Oil Field Company, Karamay, China
2 Petroleum School, China University of Petroleum (Beijing) at Karamay, Karamay, China
* Corresponding Author: Hu Yang. Email:
(This article belongs to the Special Issue: Geomechanical Issures in the Development of Reservoirs and New Energy)

Energy Engineering https://doi.org/10.32604/ee.2026.077889

Received 18 December 2025; Accepted 24 March 2026; Published online 30 April 2026

Abstract

Accurate detection of formation pore pressure is a critical technology for ensuring the safety of oil and gas drilling and reservoir evaluation. Due to the heterogeneity of geological structures and the nonlinearity of stress-strain relationships, traditional empirical formulas and numerical simulation methods often fail to meet the precision requirements when detecting formation pore pressure, constrained by the limited amount of logging data. The physically-constrained convolutional neural network-Transformer (Phys-CNN-Transformer) hybrid architecture proposed in this paper enhances the accuracy and generalization capability of formation pore pressure detection. The convolutional operations in CNN learn the mapping relationships of local features in well-logging parameters, effectively capturing local correlations within well-logging curves. The Transformer model establishes global temporal and spatial connections among eight formation parameters—Depth, acoustic interval transit time (AC), bulk density (DEN), gamma ray (GR), compensated neutron log (CNL), borehole diameter (CAL), flushed zone resistivity (RXO), and spontaneous potential (SP)—across different formation depths, thereby improving the generalization ability of the pore pressure detection model. Further physical function constraints were applied to the CNN-Transformer model: utilizing the Ldata primary loss function and Eaton’s method, which achieved an R² result of 0.968. This study utilized actual logging data from eight wells in the Mahu Oilfield (Well Mahu 1, Well Mahu 2, Well Mahu 8, Well Mahu 6, Well Mahu 15, Well Mahu 16, Well Mahu 25, and Well Mahu 26) as training and testing sets. During the training process, hyperparameter tuning and ablation comparison tests were employed to enhance the accuracy of the physically constrained CNN-Transformer model. The Phys-CNN-Transformer model was compared with other convolutional neural network models, including LSTM and BP. The results demonstrate that this approach offers a new perspective for advancing logging technology and improving the detection accuracy of formation pore pressure.

---

Keywords

Convolutional neural network; transformer; physical constraint function; Mahu oilfield; formation pore pressure detection

---