A quadratic neural network for multiphase flow prediction in CO2 sequestration

Abstract

Geological carbon sequestration (GCS) is widely recognized as a key strategy for mitigating global climate change by reducing atmospheric CO 2 concentrations. Deep learning methods, such as long short-term memory (LSTM) networks and multilayer perceptron (MLP) models, have been widely applied to multiphase flow prediction in CO 2 sequestration. However, most existing deep learning approaches struggle to capture complex nonlinear dynamics, making it challenging to achieve both high prediction accuracy and computational efficiency in large-scale scenarios. To address this limitation, a quadratic multi-layer perceptron (Quadratic-MLP) is proposed by introducing a quadratic polynomial input–output mapping at the neuron level, which enhances nonlinear representation capability while maintaining a compact network architecture. The proposed Quadratic-MLP is applied to multiphase flow prediction in CO 2 sequestration using 27 numerical simulation cases generated from three heterogeneous geological models, with a spatial resolution of 25×25×3 over 72 time steps. Results demonstrate that the Quadratic-MLP outperforms conventional LSTM and MLP models. Compared with the conventional MLP model, the proposed Quadratic-MLP reduces the mean squared error by more than 60% for all predicted variables, including pressure, saturation, and water production rate. Its prediction accuracy also surpasses that of deeper and wider MLP models with more trainable parameters, demonstrating improved parameter efficiency. Furthermore, the capability of the proposed Quadratic-MLP to capture temporal evolution trends beyond the training period highlights its potential for large-scale reservoir simulation and rapid prediction applications.