Reliable and explainable machine-learning methods for accelerated material discovery

作     者：Bhavya Kailkhura Brian Gallagher Sookyung Kim Anna Hiszpanski T.Yong-Jin Han 

作者机构：Center for Applied Scientific ComputingComputing DirectorateLawrence Livermore National LaboratoryLivermoreCAUSA Materials Science DivisionPhysical and Life Sciences DirectorateLawrence Livermore National LaboratoryLivermoreCAUSA 

出 版 物：《npj Computational Materials》 (计算材料学（英文）)

年 卷 期：2019年第5卷第1期

页      面：186-194页

核心收录：

学科分类：0502[文学-外国语言文学] 050201[文学-英语语言文学] 05[文学] 

基　　金：This work was performed under the auspices of the U.S.Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344 was supported by the LLNL-LDRD Program under Project No.16-ERD-019 and 19-SI-001(LLNL-JRNL-764864) 

主　　题：explain overcome simpler 

摘      要：Despite ML’s impressive performance in commercial applications,several unique challenges exist when applying ML in materials science *** such a context,the contributions of this work are ***,we identify common pitfalls of existing ML techniques when learning from underrepresented/imbalanced material ***,we show that with imbalanced data,standard methods for assessing quality of ML models break down and lead to misleading ***,we find that the model’s own confidence score cannot be trusted and model introspection methods(using simpler models)do not help as they result in loss of predictive performance(reliability-explainability trade-off).Second,to overcome these challenges,we propose a general-purpose explainable and reliable machine-learning ***,we propose a generic pipeline that employs an ensemble of simpler models to reliably predict material *** also propose a transfer learning technique and show that the performance loss due to models’simplicity can be overcome by exploiting correlations among different material properties.A new evaluation metric and a trust score to better quantify the confidence in the predictions are also *** improve the interpretability,we add a rationale generator component to our framework which provides both model-level and decision-level ***,we demonstrate the versatility of our technique on two applications:(1)predicting properties of crystalline compounds and(2)identifying potentially stable solar cell *** also point to some outstanding issues yet to be resolved for a successful application of ML in material science.