Prediction-accuracy improvement of neural network to ferromagnetic multilayers by Gaussian data augmentation and ensemble learning
Prediction-accuracy improvement of neural network to ferromagnetic multilayers by Gaussian data augmentation and ensemble learning
复制标题
DOI：
10.1016/j.commatsci.2023.112032
复制
发表时间：
2023-02
影响因子：
3.3
通讯作者：
K. Nawa;K. Hagiwara;Kohji Nakamura
中科院分区：
文献类型：
--
作者：
K. Nawa;K. Hagiwara;Kohji Nakamura
在使用机器学习和密度泛函理论（DFT）计算的材料信息学中，由于DFT的成本非常大，通常很难获得足够的数据库。因此，需要一种从有限的数据集中学习复杂目标关系的机器学习技术。在目前的工作中，为了克服这个问题，我们通过实现两种技术建立了一个神经网络：高斯数据增强（GDA）方法，它将高斯噪声注入训练数据集，以及集成学习，它采用多个模型来训练并通过平均其输出进行预测。与典型的例子磁矩和形成能作为原子层配置的函数在CoFe多层膜，预测精度可以大大提高，例如，通过使用由所有数据的10 - 30%组成的训练数据集。我们发现，使用GDA大大提高了未知测试数据集的预测精度，其中改进归因于NN的拟合曲线的平滑效果，并且与集成学习的组合带来了进一步的改进，除了类似的平滑效果之外，还减少了源于训练样本数据集的选择的方差误差。因此，本方法可以广泛地推广到数据库有限的材料信息学。
In materials informatics using machine learning and density functional theory (DFT) calculations, it is often hard to obtain enough database due to extremely large costs of DFT. Therefore, it is required a machine learning technique that learns a complex target relationship from a limited dataset. In the present work, to overcome this issue, we built a neural network by implementing two techniques: Gaussian data augmentation (GDA) method, which injects Gaussian noises into the training dataset, and ensemble learning, which employs multiple models to train and make prediction by averaging their outputs. With typical examples of magnetic moment and formation energy as a function of atomic-layer configuration in CoFe multilayers, the prediction accuracy can be greatly improved, eg, by using a training dataset consisting of 10∼ 30% of all data. We found that the use of GDA substantially increases the prediction accuracy for unknown test dataset where the improvement is attributed to a smoothing effect of a fitting curve of NN, and a combination with the ensemble learning brings further improvement with reducing the variance error originating from the selection of training sampling dataset in addition to a similar smoothing effect. The present approach, thus, can be generalized widely to materials informatics for which database is limited.