Fracture zone prediction using a semi-supervised neural network with limited well-labelled and extensive unlabelled seismic data

Abstract Natural fractures play an important role in controlling fluid flow and production performance, making accurate fracture zone prediction essential for petroleum exploration and development. However, prediction of fracture zones using seismic data remains difficult because fracture-related seismic responses are weak and labelled samples from wells are limited. To overcome these limitations, this study proposes a semi-supervised ladder neural network (SSLNN) for fracture zone prediction in the Asmari Formation of the A Oilfield, Iraq. Fracture development was identified from well logs, and fracture-sensitive seismic attributes were screened using support vector machine (SVM)-based sensitivity analysis. Five attributes, namely variance, dip angle, curvature, azimuth angle, and dip deviation, were selected as the optimal input features. By integrating limited labelled data with abundant unlabelled seismic attribute data, the SSLNN effectively captures the nonlinear relationship between seismic responses and fracture development. The results show that the proposed method achieves an average test accuracy of 87.5%, exceeding that of the best-performing supervised model by 6.04%. Parameter analysis further indicates that model accuracy first increases, then decreases, and finally stabilises as the weight ratio between supervised and unsupervised losses increases. The prediction results reveal that fractures are more developed in Layer A than in Layer B, are stronger in the southern part of Layer A, and are mainly concentrated near structural highs and in the hanging walls of faults. These results demonstrate that the SSLNN is an effective method for fracture prediction in carbonate reservoirs with sparse labelled data.