Using thin-section images to estimate core permeability is an economical and less time-consuming method for reservoir evaluation, which is a goal that many petroleum developers aspire to achieve. Although three-dimensional (3D) pore volumes have been successfully applied to train permeability models, it is very expensive to carry out. In this regard, deriving permeability from two-dimensional (2D) images presents a novel approach in which data are fitted directly on the basis of pore-throat characteristics extracted from more cost-effective thin sections. This work proposes a Fluid–MLP workflow for estimating 3D permeability models. We employed DIA technology combined with artificial lithology and pore classification to calculate up to 110 characteristic parameters of the pore-throat structure on the basis of 2D rock cast thin sections. The MLP network was adopted to train the permeability prediction model, utilizing these 110 parameters as input. However, the accuracy of the conventional MLP network only reached 90%. We propose data preprocessing using fluid flow simulations to improve the training accuracy of the MLP network. The fluid flow simulations involve generating a pore network model based on the 2D pore size distribution, followed by employing the lattice Boltzmann method to estimate permeability. Subsequently, six key structural parameters, including permeability calculated by LBM, pore type, lithology, two-dimensional porosity, average pore–throat ratio, and average throat diameter, were fed into the MLP network for training to form a new Fluid–MLP workflow. Comparing the results predicted using this new Fluid–MLP workflow with those of the original MLP network, we found that the Fluid–MLP network exhibited superior predictive performance.

The rapid and accurate forecasting of two-phase flow in porous media is a critical challenge in oil field development, exerting a substantial impact on optimization and decision-making processes. Although the Convolutional Long Short-Term Memory (ConvLSTM) network effectively captures spatiotemporal dynamics, its generalization in predicting complex engineering problems remains limited. Similarly, although the Fourier Neural Operator (FNO) demonstrates adeptness at learning operators for solving partial differential equations (PDEs), it struggles with three-dimensional, long-term prediction. In response to these limitations, we introduce an innovative hybrid model, the Convolutional Long Short-Term Memory-Fourier Neural Operator (CL-FNO), specifically designed for the long-term prediction of three-dimensional two-phase flows. This model integrates a 3D convolutional encoder–decoder structure to extract and generate hierarchical spatial features of the flow fields. It incorporates physical constraints to enhance the model’s forecasts with robustness through the infusion of prior knowledge. Additionally, a temporal function, constructed using gated memory-forgetting mechanisms, augments the model’s capacity to analyze time series data. The efficacy and practicality of the CL-FNO model are validated using a synthetic three-dimensional case study and application to an actual reservoir model.

Reservoir injection-production optimization plays a critical role in enhancing recovery in high-water-cut oilfields. While polymer flooding improves development efficiency by increasing viscosity and controlling fluid flow, its optimization faces several challenges: Conventional numerical simulations are computationally expensive, often requiring several hours per run, which severely limits scalability. The incorporation of slug measures introduces high-dimensional coupled modeling difficulties, as discrete variables (e.g., polymer injection timing) must be integrated with continuous variables (e.g., concentration and injection rate). Moreover, single-stage optimization frameworks are prone to converge to local optima and often fail to adapt to the dynamic characteristics of heterogeneous reservoirs. To address these issues, this study proposes a systematic solution framework: (1) A hybrid variable modeling approach is developed, which employs binary encoding to represent slug timing and a radial basis function (RBF) surrogate model based on Gower distance to map discrete-continuous variable relationships; (2) A two-stage surrogate-driven optimization architecture is introduced, where a global phase uses a dual-mode surrogate search to identify high-potential regions, followed by a local refinement phase for continuous parameter adjustment. An adaptive switching mechanism is incorporated to balance exploration and exploitation. Experimental results demonstrate that the proposed Two-Stage Collaborative Optimization (TSCO) framework outperforms state-of-the-art Surrogate-Assisted Evolutionary Algorithms (SAEAs) on multiple benchmark problems.