Abstract The long-term potential capacity for soils to deliver essential elements to plants and the soil solution is often assessed using aqua regia (AR) extractions, and understanding specific mineral contributions to extracted element concentrations is central to the accurate interpretation of such data. The objective of this study was to assess whether the supplying quantitative soil mineralogy data to the Cubist machine learning algorithm could result in accurate prediction of AR extractable base cation concentrations (Naaqr, Mgaqr, Kaqr and Caaqr), and consequently facilitate the interpretation of meaningful mineral sources of these elements. Mineral concentrations were quantified in 191 soil samples collected by horizon from 96 sites across Scotland. These mineral concentrations together with pH and soil organic matter data were used to train Cubist models for the prediction of log(Naaqr), log(Mgaqr), log(Kaqr), and log(Caaqr), yielding concordance correlation coefficients of 0.77, 0.87, 0.78 and 0.71, respectively when assessed by cross validation. The dominant variables selected by Cubist for the conditions of the models contrasted substantially between the four elements, aligning with the independent nature of each base cation interpreted from compositional biplot analysis. Namely, plagioclase concentrations were predominantly selected by Cubist for log(Naaqr) prediction, chlorite concentrations for log(Mgaqr) prediction, dioctahedral mica concentrations for log(Kaqr) prediction, and pH for log(Caaqr) prediction. The selection of these variables is found to be geochemically appropriate for each element, demonstrating how data-driven methods can yield meaningful conclusions with respect to mineral sources of AR extractable elements. In validating the utility of combining quantitative soil mineralogy data with the Cubist machine learning algorithm on the relatively well characterised base cations, future work can progress towards using similar approaches to define mineral sources of more challenging trace elements in new detail.

中文翻译：

---

从使用定量矿物学数据训练的 Cubist 模型解释的苏格兰土壤中王水可提取碱阳离子的矿物来源

摘要 土壤向植物和土壤溶液输送必需元素的长期潜在能力通常使用王水 (AR) 提取进行评估，了解特定矿物质对提取元素浓度的贡献对于准确解释此类数据至关重要。本研究的目的是评估向 Cubist 机器学习算法提供定量土壤矿物学数据是否可以准确预测 AR 可提取的碱性阳离子浓度（Naaqr、Mgaqr、Kaqr 和 Caaqr），从而促进对有意义矿物的解释这些元素的来源。通过地平线从苏格兰 96 个地点收集的 191 个土壤样品中的矿物质浓度进行了量化。这些矿物质浓度与 pH 值和土壤有机质数据一起用于训练 Cubist 模型，以预测 log(Naaqr)、log(Mgaqr)、log(Kaqr) 和 log(Caaqr)，产生 0.77、0.87 的一致性相关系数当通过交叉验证评估时，分别为 0.78 和 0.71。Cubist 为模型条件选择的主要变量在四个元素之间形成鲜明对比，与从成分双标分析中解释的每个碱基阳离子的独立性质一致。即，斜长石浓度主要由 Cubist 选择用于 log(Naaqr) 预测，绿泥石浓度用于 log(Mgaqr) 预测，二八面体云母浓度用于 log(Kaqr) 预测，以及 pH 用于 log(Caaqr) 预测。发现这些变量的选择在地球化学上适用于每个元素，证明了数据驱动的方法如何能够产生关于 AR 可提取元素的矿物来源的有意义的结论。在验证将定量土壤矿物学数据与 Cubist 机器学习算法对相对较好表征的碱基阳离子相结合的效用时，未来的工作可以朝着使用类似方法以新的细节定义更具挑战性的微量元素的矿物来源的方向发展。