Magnesium potassium phosphate cement (MKPC) is a kind of Mg-chemically bonded phosphate ceramic commonly used for rapidly repairing dilapidated structures. In this study, a compressive strength dataset of MKPC was constructed, and four advanced machine learning (ML) algorithms (XGBoost, RF, GBDT and ANN) were selected to establish a high-precision compressive strength prediction model of MKPC. The SHAP and PDP methods are also used for interpretability analysis of ML-MKPC models. The XGBoost model has good generalizability and reliability while achieving high prediction accuracy. The RF and GBDT models performed similarly to the XGBoost model on the training set but performed poorly on the testing set. The ANN model is poorly trained on both the training and testing sets, with a risk of underfitting. The R of the XGBoost model at the different compressive strength stages still reaches above 0.80, indicating that it not only captures the complex relationships of the overall dataset well but also effectively predicts the staged strength dataset. Feature importance analysis revealed that the curing age (T), water-to-binder ratio (W/B), mineral admixtures-to-binder ratio (MA/B) and phosphate-to-magnesium ratio (P/M) are the principal variables affecting the compressive strength of MKPC. The partial interpretation shows that the optimum value range is determined when W/B is 0.10–0.18, MA/B is 0–0.20, P/M is 0.40–1.0, and R/M is 0–0.12. The composition of mineral admixtures with high-Ca, high-Si and low-Al systems seems to be more conducive to participating in the hydration reaction of MKPC. The ML-MKPC compressive strength prediction model developed in this study can provide theoretical support for the subsequent composition design and performance optimization of MKPC.

中文翻译：

---

基于可解释的机器学习方法深入洞察MKPC抗压强度发展的驱动因素


磷酸镁钾水泥（MKPC）是一种镁化学结合磷酸盐陶瓷，常用于快速修复破损结构。本研究构建了MKPC的抗压强度数据集，并选择了四种先进的机器学习（ML）算法（XGBoost、RF、GBDT和ANN）来建立MKPC的高精度抗压强度预测模型。 SHAP 和 PDP 方法也用于 ML-MKPC 模型的可解释性分析。 XGBoost模型具有良好的泛化性和可靠性，同时实现了较高的预测精度。 RF 和 GBDT 模型在训练集上的表现与 XGBoost 模型类似，但在测试集上表现不佳。 ANN 模型在训练集和测试集上的训练都很差，存在欠拟合的风险。 XGBoost模型在不同抗压强度阶段的R仍然达到0.80以上，表明它不仅很好地捕捉了整体数据集的复杂关系，而且可以有效地预测阶段强度数据集。特征重要性分析表明，养护龄期（T）、水胶比（W/B）、矿物掺合料与粘结剂比（MA/B）和磷酸盐与镁比（P/M）是影响 MKPC 抗压强度的主要变量。部分解释表明，最佳取值范围确定为W/B为0.10~0.18、MA/B为0~0.20、P/M为0.40~1.0、R/M为0~0.12。高钙、高硅、低铝体系的矿物掺合料组成似乎更有利于参与MKPC的水化反应。本研究开发的ML-MKPC抗压强度预测模型可为后续MKPC的成分设计和性能优化提供理论支持。