融合材料领域知识的数据准确性检测方法

Detection Method on Data Accuracy Incorporating Materials Domain Knowledge

材料数据由于小样本、高维度、噪音大等特性,用于机器学习建模时常常会产生与领域专家认知不一致的结果。面向机器学习全流程,开发材料领域知识嵌入的机器学习模型是解决这一问题的有效途径。材料数据的准确性直接影响了数据驱动的材料性能预测的可靠性。本研究针对机器学习应用过程中的数据预处理阶段,提出了融合材料领域知识的数据准确性检测方法。该方法首先结合材料专家认知构建了材料领域知识库。然后,将其与数据驱动的数据准确性检测方法结合,从数据和领域知识两个角度对材料数据集进行基于描述符取值规则的单维度数据正确性检测、基于描述符相关性规则的多维度数据相关性检测以及基于多维相似样本识别策略的全维度数据可靠性检测。对于每一阶段识别出的异常数据,结合材料领域知识进行修正,并将领域知识融入到数据准确性检测方法的全过程以确保数据集从初始阶段就具有较高准确性。最后该方法在NASICON型固态电解质激活能预测数据集上的实验结果表明:本研究提出的方法可以有效识别异常数据并进行合理修正。与原始数据集相比,基于修正数据集的6种机器学习模型的预测精度都有不同程度的提升。其中,在最优模型上R2提升了33%。

Due to the characteristics of small samples, high dimensions, and much noise, materials data often produce inconsistent results with those obtained from domain experts when used for machine learning modeling.For the whole process of machine learning, developing machine learning models embedding materials domain knowledge is a solution to this problem. The accuracy of materials data directly affects the reliability of data-driven materials performance prediction. Here, a data accuracy detection method incorporating materials domain knowledge is proposed by focusing on the data preprocessing stage in the machine learning application process. Firstly, a materials domain knowledge database is constructed based on the knowledge from materials experts. Secondly, it is coordinated with the data-driven data accuracy detection method to perform single-dimensional data accuracy detection based on the rule for value of descriptors, multi-dimensional data correlation detection based on the rule for correlation of descriptors, and full-dimensional data reliable detection based on multi-dimensional similar sample identification strategy from both data and domain knowledge perspectives. For the anomalous data identified at each stage, they are corrected by incorporating the materials domain knowledge. Furthermore, domain knowledge is incorporated into the whole process of the data accuracy detection method to ensure high accuracy of the dataset from the initial stage. Finally, experiments on the NASICON-type solid electrolyte activation energy prediction dataset demonstrate that this method can effectively identify anomalous data and make reasonable corrections. Compared with the original dataset, the prediction accuracy of all six machine learning models based on the revised dataset is improved to different degrees, among which R~2 achieves a 33% improvement on the optimal model.