Abstract:

Waste drilling fluids are difficult to reuse because they are strongly alkaline and remain highly stable as dispersed colloidal systems, which hinders solid-liquid separation. Therefore, gel breaking and destabilization are required before reuse. The destabilization efficiency is controlled by multiple operating conditions, among which breaker dosage, reaction temperature, pH regulator dosage, and inorganic salt proportion are critical. In this study, the Zeta potential of the treated drilling fluid was used as the stability indicator because it sensitively reflected changes in particle aggregation and dispersion. Based on experimental data obtained from a Box-Behnken design (BBD), response surface methodology (RSM) combined with analysis of variance (ANOVA) was employed to assess the significance of main effects and interactions among factors. Building on these results, three surrogate models—a backpropagation neural network (BP-ANN), random forest regression (RF), and Gaussian process regression (GPR)—were constructed to predict and compare the performance of Zeta potential modeling. Results indicated that the BP neural network achieved optimal fitting performance, with an overall coefficient of determination (R²) of 0.9629 and an average R² of 0.8125 under 5-fold cross-validation—significantly higher than those of other surrogate models. Thus, it was selected as the process optimization model. To further identify the optimal solution, a genetic algorithm (GA) was integrated with BP-ANN for global optimization. The GA-BP method yielded optimal parameters: gel breaker dosage of 3.306 g/L, reaction temperature of 57.74℃, pH adjuster dosage of 3.45 g/L, an inorganic salt proportion of 0.80, and a predicted Zeta potential of -10.735 mV. The measured value was -10.85 mV, with a relative error of 1.06%. Overall, the proposed experimental design-based modeling and optimization framework provides a reliable basis for determining operating parameters for gel breaking and destabilization, and supports the subsequent reuse of treated drilling fluids.

Key words: gel breaking, Zeta potential, Box-Behnken design, BP neural network, genetic algorithm, 5-fold cross-validation

摘要：

为实现钻井废液的高效回用，以Zeta电位表征体系稳定性，研究废弃钻井液破胶过程中关键工艺因素对稳定性的影响。首先采用单因素实验系统考察破胶剂用量、反应温度、pH调节剂加注量以及破胶剂中无机盐占比对Zeta电位的影响规律。随后基于Box-Behnken设计构建响应面实验体系，并利用响应面法对关键因素交互作用进行分析，为参数优化提供可解释依据。依据实验数据分别建立BP神经网络、随机森林(RF)回归与高斯过程回归(GPR)三种代理模型，并结合5折交叉验证对模型预测性能与泛化能力进行对比评价。研究表明，BP模型综合表现最优，其整体决定系数R²为0.9629，且5折交叉验证平均决定系数R²为0.8125，显著优于其他代理模型，能较为精准地预测Zeta电位。基于BP神经网络并结合遗传算法(GA)进行全局寻优，得到的最佳工艺为：破胶剂用量3.306 g/L、反应温度57.74℃、pH调节剂加注量3.45 g/L、无机盐占比0.80，对应预测Zeta电位为-10.735 mV。该研究结果为废弃钻井液脱稳以及后续回用提供了理论依据和数据支持。

关键词: 破胶, Zeta电位, Box-Behnken设计, BP神经网络, 遗传算法, 5折交叉验证