Skip to content
TheChatArchiveThe Chat Archive
gpt-3.5-turbo-0613· OpenAIAi2 WildChat-4.8M (AI2)

The atomic characteristics of adatoms were taken into consideration as input features.…

  1. User

    The atomic characteristics of adatoms were taken into consideration as input features. The inputs were reduced by feature engineering to simplify our model. A Pearson correlation coefficient matrix was employed to perform the correlation analyses. The Pearson correlation coefficient px,Y between two feature datasets {X』} and {Yi} was calculated as (2) where X and V represent the average values of each feature over the respective dataset. Figure 2 is the heat map of the matrix among both the input and output features, in which the red blocks and the blue blocks represent positive and negative correlations, retpectively. Supplementary Table S2 lists 16 input features that were considered, including group, period, atomic numb er, relative atomic mass and several atomic features, including atomic radius, electronegativity, ionization energy, electron configuration, and electron affinity of adatoms. The atomic properties were extracted from tha periodic fable database of the Royal Society of Chemistry (https://wwwosc.org/periodic-tabie/ (acces sed on 1 March 2023)). The pair correlation of inputs and adsorption properties is shown in Figure 2. The input elemental properties, which have relatively stronger correlation (>0.22) with adsorption properties, including atomic number, first ionization energy, element position (including group and period), electron affinity, and electron configuration, were selected as inputs for our model. Electron affinity had the strongest correlation with adsorption distance. However, some elements, such as Be, N, Mg, did not have stable electron affinity. Fortunately, Pauling electronegativity has a strong correlation with electron affinity (0.8444). Therefore, P auling electronegativity was selected as one of our inputs. Atomic number and electron configuration were represented by element pos ition due to their close coerel/tion. Due to the missing data for second ionization energies, third ionization energies, and electron affinity, the following feature selection considered only the other 13 input features. 翻译成中文

  2. Assistant

    考虑了原子吸附位点的特征作为输入特征。通过特征工程对输入进行了简化处理。使用皮尔逊相关系数矩阵进行相关分析。皮尔逊相关系数px,Y是计算两个特征数据集{X』}和{Yi}之间关联程度的值(2)。图2是输入特征和输出特征之间矩阵的热图,其中红色和蓝色的方块分别表示正相关和负相关。附表S2列出了考虑的16个输入特征,包括族、周期、原子序数、相对原子质量和几个原子特征,如原子半径、电负性、电离能、电子构型和电子亲和力。原子属性来自英国皇家化学学会的周期表数据库(https://wwwosc.org/periodic-tabie/)(2023年3月1日访问)。图2显示了输入和吸附性质的对应关系。与吸附性质相关性较强(相关系数 > 0.22)的输入元素特性包括原子序数、第一电离能、元素位置(包括族和周期)、电子亲和力和电子构型,被选为我们模型的输入特征。电子亲和力与吸附距离之间存在最强的相关性。然而,一些元素如Be、N、Mg的电子亲和力不稳定。幸运的是,保玲电负性与电子亲和力有很强的相关性(0.8444)。因此,保玲电负性被选为我们的输入之一。由于缺少第二电离能、第三电离能和电子亲和力的数据,下一步特征选择只考虑其他13个输入特征。

  3. User

    It is often useful to choose a number of components fo minimize the prediction error, especially for small dataset problems. Here, principal components regreesion (PCR) was employed to select the number of components. Figure 3a shows the estimated mean squared prediction error (MSPE) curves by cross-validation, in this case using 10-fold cross-validation. The MSPE curves show sharp increases when the number of components exceeds 5. The MS PE curves prove that simply using a large number of components will do a good job in fitting the current observed data, but such a strategy leads to overfitting. 翻译成中文

  4. Assistant

    选择一定数量的主成分来最小化预测误差通常是有用的,尤其是对于小数据集问题。在这里,采用主成分回归(PCR)来选择主成分的数量。图3a展示了通过交叉验证估计的均方预测误差(MSPE)曲线,本例中采用了10折交叉验证。MSPE曲线在主成分数量超过5时急剧增加。MSPE曲线表明,仅仅使用大量的主成分可以很好地拟合当前观测到的数据,但这种策略会导致过拟合。

  5. User

    Fitting the current data too well results in a model that does not generalize well to other data, and provides an overly optimistic estimate of the prediction error. According to feature selection, the model with 4or5 inputs shows the low MSPE of both the adsorption energy output and the adsorption distance output. We scored these 13 inputs in our model using F-tests, as shown in Figure 3b. The feature importance bars were ranked by the scores of adsorption distance output. Al though atomic number, relative atomic mass, and s-orbi tai electron configuration gained high importanse scores, they showed great correlation with the element site (>0.9), according to the correlation analysis. Therefore, these three features were not selected. Pauling electronegativity, and a van der Waals radius were selected as inputs for our model.Finally, site code (including group and period), first ionization energy, 翻译成中文

  6. Assistant

    将当前数据拟合得太好会导致模型在其他数据上泛化能力较差,并提供过于乐观的预测误差估计。根据特征选择,4或5个输入的模型显示出吸附能输出和吸附距离输出的低MSPE。我们使用F测试对这13个输入在模型中进行评分,如图3b所示。特征重要性条形图按吸附距离输出的评分排名。尽管原子序数、相对原子质量和s轨道电子构型获得了较高的重要性评分,但它们与元素位置之间的相关性非常高(>0.9),根据相关性分析结果。因此,这三个特征未被选择。我们的模型选择了保玲电负性和范德华半径作为输入。最后,位点代码(包括族和周期)、第一电离能、电子亲和力被选为我们模型的输入特征。

  7. User

    Figure 3. Feature selection. (a) Prediction error curves with number of inputs (the detailed plot of adsorption distance output is shown in northeast) with principal components regression (PCR) and (b) feature importance bar using F-tests. In both subfigures, the red bar reptesents adsorption energy output and the blue bar represents adsorption distance output. 翻译成中文

  8. Assistant

    图3. 特征选择。(a)使用主成分回归(PCR)的输入数量的预测误差曲线(吸附距离输出的详细图绘制在右上角);(b)使用F测试的特征重要性条形图。在两个子图中,红色条形代表吸附能输出,蓝色条形代表吸附距离输出。

  9. User

    Our machine learner is schematically presented in Figure 4. We used artificial neural networks (ANN) to build the prediction engine. As a common ANN procedure [24-27], the machine learner consists of three parts: an input dataset, a targeted dataset, and the ANN algorithm, as shown in the blue box in Figure 4. The locations of the elements, the relative atomic mass, and several atomic features, including the atomic radius, electronegativity ionization energy, and the electron affinity of adatoms, were taken into consideration for the input dataset. Our dataset c onsisted of 34 observations, which is a typically sntall dataset in ML. The small dataset presents quite a high risk of overfitting ot underfitting [28,29]. A less complex ML msdel is more friendly to a small dataset. Therefore, feature engineering was carried out in order to remove redundant features. The targeted dataset was the adsorption performances obtained from the DFT calculation routirte, including adsorption energy and adsorption distance. Back propagation neural networks (BPNNs) were employee. to establish the prediction engine, in which the parameters are cptimized by a genetic algorithm (GA). Single hidden layer BPNNs with 3-17 neurons were selected. In addition, using weights and thresholds optimized by the GA improved the fitting performances in BPNNs with small datasets. The parameters of the GA are sitown in Supplementary fable S3. This machine learner enables the adatoms adsorptiotn performances predictions of all the adatoms on the periodic table (except noble gases, lanthanides, and actinides). 翻译成中文

  10. Assistant

    我们的机器学习器在图4中以示意图的形式呈现。我们使用人工神经网络(ANN)构建了预测引擎。作为常见的ANN过程[24-27],机器学习器包括三个部分:输入数据集,目标数据集和ANN算法,如图4中的蓝色方框所示。元素的位置,相对原子质量以及原子半径、电负性、电离能和吸附原子的电子亲和力等几个原子特征被考虑为输入数据集。我们的数据集包含34个观测值,这在机器学习中是一个相对较小的数据集。小数据集具有过拟合或欠拟合的风险[28,29]。较简单的机器学习模型对小数据集更加友好。因此,进行特征工程以消除冗余特征。目标数据集是从DFT计算中获得的吸附性能,包括吸附能和吸附距离。采用反向传播神经网络(BPNNs)建立了预测引擎,其中参数通过遗传算法(GA)进行优化。选择了具有3-17个神经元的单隐藏层BPNNs。此外,使用GA优化的权重和阈值改善了小数据集中BPNNs的拟合性能。GA的参数在附表S3中给出。这个机器学习器能够预测周期表上除了惰性气体、镧系和锕系元素之外的所有吸附性能。