我们在电化学储能应用中使用了IEMAP,以展示该平台的能力。我们专注于为钠离子电池(SIBs)设计增强型层状氧化物正极,钠离子电池被认为是锂离子电池的有前途的替代品,因为钠储量丰富且成本低廉,并且可以重复利用大部分现有的锂离子电池(LIBs)制造基础设施[4,7,8,9]。P2型NaMnO2并且其衍生物因其在适当设计下结合了相对较高的工作电压、极具竞争力的容量和良好的循环稳定性,而引起了相当大的关注[26]. 提升层状氧化物作为正极材料性能的一种策略是利用高通量DFT计算,通过将基体晶体中的部分过渡金属(如Mn)替换为Ni和Fe等氧化还原活性阳离子,以及Ti和Li等氧化还原惰性稳定剂,来优化氧化还原电压和扩散势垒,从而设计晶体结构。然而,过渡金属层中可能存在的成分组合、掺杂物种类和阳离子分布数量呈组合爆炸式增长,这使得即使在大型高性能计算(HPC)设施上,对设计空间进行纯量子力学探索的成本也高得令人望而却步[5]。然而,在IEMAP中实现的geoCGNN机器学习模型的使用,帮助对超过500,000种NaMnO衍生物进行了系统且快速的筛选2,在网络权重上传后,在NVIDIA V100 GPU上仅需约10−3秒即可预测单个形成能和氧化还原电位[6]。NaMnO2的结构如图所示图3,其衍生物通过将Mn原子替换为Ni和Ti原子获得,从而得到通用体系:NaMnxNiyTiz此外,IEMAP通过geoCGNN推理服务与MongoDB数据库之间的交互,将部分Mn-Ni-Ti成分的形成能和平均电压的DFT计算结果存储其中,从而实现了对机器学习预测的DFT修正,这有助于根据稳定性和电压筛选和选择我们通用系统中Mn、Ni和Ti的最佳含量。更多晶体学信息和机器学习性能的定量细节见[2. Furthermore, IEMAP, through the interplay between the geoCGNN inference service and the MongoDB database, where the results of DFT calculations of formation energy and average voltage for some Mn-Ni-Ti compositions were stored, enabled DFT corrections on the ML predictions that facilitated the screening and selection of the best amounts of Mn, Ni, and Ti in our general system with respect to stability and voltage. Further crystallographic information and the quantitative details on the ML performances are provided in [6]. 图3。 () P2层状氧化物NaMnO的原子模型超胞的(2)侧视图和()顶视图。蓝色、紫色和红色分别代表Mn、Na和O原子。) side and (c) top views. In blue, purple, and red are the Mn, Na, and O atoms, respectively.
4.2. 参考DFT数据集与IEMAP摄取
在此背景下,建立了一套工作流程,利用密度泛函理论(DFT)计算形成能和氧化还原电位f针对一组精选的Na–Mn–Ni–Ti成分及阳离子排列,这些成分和排列源自P2-NaMnO基晶体 [redox图32前者提供了与材料结构稳定性及循环寿命相关的指标;后者则与使材料在充放电状态间转换的嵌入反应相关,并与能量密度相关。生成的数据库通过iemap-mi Python客户端导入IEMAP,该客户端确保每种配置都由一致的元数据描述,包括材料化学式、结构原型、计算方法、收敛参数和目标属性,以及完整的溯源信息[6] (]。这些精选数据既可作为后续分析的参考语料库,也可作为平台中机器学习模型的训练和验证输入。). The former gives indications related to the structural stability of the composition and then to the cycle life; the latter is associated with the intercalation reaction that brings a discharged material to the charged one or vice versa, and it is related to energy density. The resulting dataset is ingested into IEMAP through the iemap-mi Python client, which ensures that each configuration is described by consistent metadata specifying material formula, structural prototype, computational method, convergence parameters and target properties, together with full provenance information [27]. These curated data serve both as a reference corpus for subsequent analyses and as training and validation input for machine-learning models deployed in the platform.
为了评估机器学习方法,选择了一组具有不同锰、钛和镍原子组合的代表性构型。在图4中,图4f (a) 和的预测值图4b) 与DFT方法计算并托管在IEMAP数据集中的相应值进行比较。比较结果表明预测值与DFT计算值之间存在令人满意的吻合度,特别是在形成能方面,我们得到的MAE = 0.041 eV,RMSE = 0.062 eV,而对于redox (我们得到的MAE = 0.465 V,RMSE = 0.617 V。在这两种情况下,我们都观察到数据点沿x = y参考线(由红色虚线表示)对齐。相比于redox,其准确度更高这是由于两个数据集的规模不同。MP数据库包含超过15万个形成能数值,但氧化还原电位的数值则少两个数量级(不足4500个)。在更大的数据集上进行训练,对于监督式机器学习方法能产生更准确的预测,这一点从前者所考察的物理量比后者具有更小的误差条中得到了证实。误差条是通过采用k折交叉验证技术获得的,其中k=15,如参考文献[所示。],在训练期间。这为网络提供了15组权重,因此对于给定的量可以得到15个可能的预测值。通过这种方式,用户可以估计平均值和f,并将其与DFT计算结果进行比较,例如通过参考文献[]中描述的修订预测器定义,以检查geoCGNN是否返回准确的预测值并进行适当的修正,这对于V特别有益redox is due to the different sizes of the two datasets. The MP database contains more than 150 thousand values for formation energy but two orders of magnitude fewer values for the redox potential (less than 4500 values). Training on a larger dataset yields more accurate predictions for supervised ML methods, as confirmed by the smaller error bars for the former physical quantity inspected compared to the latter. The error bars are obtained by adopting the k-fold cross validation technique with k = 15 as indicated in ref. [6], during training. This provides 15 sets of weights for the networks and hence 15 possible predictions for a given quantity. In this way, the user can estimate mean value for Ef and Vredox which can be compared with the DFT calculations, through for instance the definition of a revised predictor as described in ref. [6], to check whether geoCGNN returns accurate predictions and to perform the proper correction, which is particularly beneficial for Vredox. 图4。形成能),绿色符号)和氧化还原电位f ((),蓝色符号)预测值与DFT计算值,针对具有不同Mn、Ti和Ni原子组合的选定构型结构。图表改编自参考文献[), green symbols) and redox potential Vredox ((b), blue symbols) predictions values vs. DFT calculated values for structures of selected configurations with different combinations of Mn, Ti and Ni atoms. Figures are adapted from ref. [6]. 在通过DFT计算验证了机器学习模型后,我们将其应用扩展到Mn-Ni-Ti组合的整个成分空间,预测每种成分的p形成能值Ef和氧化还原电位Vredox。由于此类材料的理想候选者应结合上述两种特性,即循环性(低Ef)和能量密度(高Vredox),一个多目标函数MO被定义为整合这两个物理量。 MO (p) = −Ef/e + Vredox