An interpretable machine learning strategy for pursuing high piezoelectric coefficients in (K_0.5Na_0.5)NbO₃-based ceramics 

Bowen Ma, Xiao Wu, Chunlin Zhao, Cong Lin, Min Gao, Baisheng Sa & Zhimei Sun

npj Computational Materials 9: 229 (2023); Published online: 22 December 2023

Editorial Summary

Machine learning for interpretable KNN ceramic high piezoelectric coefficients: fast and good?

People have invested a lot of time and energy in making piezoelectric ceramics lead-free. Through cumbersome and complicated composition control, the piezoelectric coefficient of KNN-based ceramics has been breaking through year by year. On the one hand, the traditional empirical trial-and-error research paradigm has the limitation of time-consuming consumables. On the other hand, the literature data related to the piezoelectric coefficient (d₃₃) of KNN-based ceramics accumulated in the past is extremely valuable to explore. This study constructed a d₃₃ descriptor development framework and proposed a descriptor containing only four easily accessible parameters: is used to predict d₃₃. This descriptor can also explain the high d₃₃ mechanism caused by the coexistence of multi-phase KNN-based ceramics. The team of Prof. Wu Xiao and Prof. Sa Baisheng from Fuzhou University and Prof. Sun Zhimei from Beihang University established a regression mapping of the chemical composition of KNN-based ceramics with d₃₃ from 1113 data points in 244 published articles by coupling feature engineering, machine learning regression and the SISSO algorithm. This study constructed global and local features based on the element position configuration of ABO₃-type perovskites, and then used methods such as Pearson correlation screening, feature importance, and feature exhaustion to screen key features. The optimal extreme random tree regression model has a leave-one-out cross-validation error as low as ±49pC/N.The author used the optimal feature set for the SISSO descriptor search and obtained a descriptor that has an intuitive changing trend with d₃₃.At the same time, the descriptor value also has a mapping relationship with the KNN-based ceramic phase boundary, that is, chemical compositions with descriptor values in a smaller range are more likely to obtain high d₃₃.This method has been verified in the newly published 63 KNN-based ceramic high piezoelectric coefficient compositions. This study established a mathematical mapping model of the d₃₃-composition-phase boundary in KNN-based ceramics, providing a highly intuitive and instructive way to improve the performance of perovskites, and overcoming the low reliability and difficult interpretation problems of traditional machine learning models. This article was recently published in npj Computational Materials 9, 229 (2023).

编辑概述

可解释KNN陶瓷高压电系数的机器学习：又快又好？

高性能压电陶瓷是现代电子设备中至关重要的一类材料。为追求可持续发展，以PbZr_1−xTi_xO₃ (d₃₃ ≈ 200-1500 pC/N)为主的传统铅基压电材料正在逐步过渡到(K, Na)NbO₃ (KNN)基等无铅压电陶瓷。经过数十年的努力，KNN基陶瓷压电系数(d₃₃)不断突破(d₃₃ ≈ 100-650 pC/N)，同时也积攒了大量的文献数据有待进一步挖掘，但基于传统经验试错法的研究范式难以拜托耗时耗力的局限性。近年来，伴随着人工智能和材料信息学的发展，机器学习在挖掘材料结构/组分-性能方面展现出强大的潜力。同时，符号回归、确定独立性筛选和稀疏化算子（SISSO）等兼具性能和可解释性的机器学习方法也应运而生，为非计算机领域的研究人员提供了更直观和易于理解的新研究途径。福州大学吴啸副教授、萨百晟教授联合北京航空航天大学孙志梅教授团队，通过耦合特征工程、机器学习回归和SISSO算法，构建了一种d₃₃描述符开发框架，提出一个仅包含4种易于获取参数的描述符：用于预测KNN基陶瓷的d₃₃。该研究从244篇已发表的1113个数据点中建立了KNN基陶瓷化学成分与d₃₃的回归映射。作者根据ABO₃型钙钛矿的元素位置构型，分别从全局和局部进行特征构造，然后利用Pearson相关性筛选、特征重要性和特征穷尽等方法对关键特征进行筛选，最优的极端随机树回归模型的留一交叉验证误差最低至±49 pC/N。最后将最优的特征集用于SISSO描述符搜索，得到了一个和d₃₃有直观变化趋势的描述符，同时该描述符数值还与KNN基陶瓷相界具备映射关系，即描述符值在较小区间的化学成分更容易获得高压电系数。这一方法在最新发表的63个KNN基陶瓷高压电系数组分中得到验证。该研究在KNN基陶瓷中建立了d₃₃-组分-相界三者的数学映射模型，为提高钙钛矿的性能提供了一种高度直观和指导性的途径，克服了传统机器学习模型低可信度和难解释的问题。