基于机器学习的超导临界温度预测

Prediction of Superconducting Critical Temperature Based on Machine Learning

过低的超导临界温度(T_(c))一直是限制超导材料广泛应用的主要原因,因此如何设计出具有较高T_(c)的合金是超导材料发展的关键步骤。本工作以机器学习算法为工具,基于两种特征生成器和九种回归算法,建立了预测T_(c)的数据驱动模型,提高了新型超导合金的研发效率。结果表明,极端随机树算法(etr)和Magpie规则的预测效果最好,决定系数(R^(2))和平均绝对误差(MAE)在测试集上分别能达到0.88和2.21 K。为降低模型的复杂程度,对生成的特征通过过滤法和穷举法进行特征筛选,再结合特征的重要性排序,得到对T_(c)影响最大的三种特征分别是电负性的平均绝对偏差、平均未填充电子数和平均s层未填充电子数。以三种特征为变量,建立T_(c)的预测图谱,发现在平均s层未填充电子数为0.2~0.6,电负性的平均绝对偏差为0.5~1.2,平均未填充电子数为0~5时,材料有着大于50 K的T_(c),为新型超导合金的成分设计提供方向。

The low superconducting critical temperature(T_(c))of has always been a major barrier to their widespread application.Therefore,designing alloys with higher T_(c) is a crucial step in the development of superconducting materials.In this work,a data⁃driven model for predicting T_(c) was established using machine learning algorithms based on two feature generators and nine regression algorithms,improving the efficiency of new superconducting alloy development.The results showed that the extreme random trees algorithm(etr)and the Magpie⁃generated rule had the best prediction performance,with a coefficient of determination(R^(2))and mean absolute error(MAE)of 0.88 and 2.21 K,respectively,on the test data set.To reduce the complexity of the model,feature selection was performed on the generated features using filter and exhaustive search methods,combined with feature importance ranking.The three features that had the greatest influence on T_(c) were the mean absolute deviation of electronegativity,the average unfilled electron count,and the average unfilled electron count of the s layer.A T_(c) prediction map was established using these three features,which indicated that materials with T_(c) greater than 50 K could be achieved when the average unfilled electron count of the s layer was between 0.2 and 0.6,the mean absolute deviation of electronegativity was between 0.5 and 1.2,and the average unfilled electron count was between 0 and 5,providing a direction for the composition design of new superconducting alloys.