文章信息/Info

Title:

Attribute Prediction of Aluminum Alloy Based on Machine Learning

作者:

左厚辰1; 江永全*2; 楊燕2

1 西南交通大學(xué)牽引動(dòng)力國(guó)家重點(diǎn)實(shí)驗(yàn)室 四川省 成都市 610031， 2 西南交通大學(xué)計(jì)算機(jī)與人工智能學(xué)院 四川省 成都市 611756

Author(s):

Houchen Zuo 1;  Yongquan Jiang *2;  Yan Yang 2

1.State key labratory of traction power, Southwest Jiaotong University, Chengdu 610031, China, 2. School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu 611756, China

關(guān)鍵詞:

合金材料; 性質(zhì)預(yù)測(cè); 人工智能;  機(jī)器學(xué)習(xí);  自動(dòng)機(jī)器學(xué)習(xí)

Keywords:

alloy material;  property prediction;  artificial intelligence;  machine learning;  automatic machine learning

文獻(xiàn)標(biāo)志碼:

A

摘要:

當(dāng)前材料領(lǐng)域?qū)τ趯傩缘挠?jì)算主體上還是以密度泛函以及它的優(yōu)化算法為主，雖然計(jì)算結(jié)果準(zhǔn)確，但需花費(fèi)大量的時(shí)間和資源。近年來(lái)人工智能在材料領(lǐng)域應(yīng)用廣泛，在材料性質(zhì)預(yù)測(cè)領(lǐng)域，不少學(xué)者引用機(jī)器學(xué)習(xí)算法進(jìn)行實(shí)驗(yàn)，取得了不錯(cuò)的效果。實(shí)驗(yàn)針對(duì)原子平均體積、原子平均能量以及原子形成能三個(gè)性質(zhì)開展，使用的數(shù)據(jù)集來(lái)自于開放量子材料數(shù)據(jù)庫(kù)OQMD，通過(guò)支持向量機(jī)模型、梯度提升回歸模型、自動(dòng)機(jī)器學(xué)習(xí)auto_ml和AutoKeras以及基于AutoKeras最佳模型改進(jìn)的深度全連接網(wǎng)絡(luò)DNN和殘差網(wǎng)絡(luò)ResNet，驗(yàn)證機(jī)器學(xué)習(xí)在材料屬性預(yù)測(cè)的可行性和準(zhǔn)確性。實(shí)驗(yàn)結(jié)果表明，基于自動(dòng)機(jī)器學(xué)習(xí)AutoKeras最佳模型改進(jìn)的ResNet效果最佳，原子平均體積的最佳模型的R-Square為0.9655，MAE為0.6306*10-30m3/atom，原子平均能量的最佳模型的R-Square為0.9724，MAE為0.1466 eV/atom，原子形成能的最佳模型的R-Square為0.9880，MAE為0.0732 eV/atom。

Abstract:

At present, the calculation of properties in the field of materials is mainly based on density functional theory and its optimization algorithm. Although the calculation results are accurate, it costs a lot of time and source. In recent years, artificial intelligence is widely used in the field of materials. In the field of material property prediction, many scholars use machine learning algorithm to experiment and achieve good results. The experiment is carried out for the three properties of average atomic volume, average atomic energy and atomic formation energy. The data set used is the open quantum material database. Through support vector machine model, gradient boosting regression model, automatic machine learning auto_ml and AutoKeras, deep network fully connected network and residual network, experiments verify the feasibility and accuracy of machine learning in material attribute prediction. The experimental results show that the improved models based on the best model provided by automatic machine learning AutoKeras have the best effect. The R-Square of the best model of atomic average volume is 0.9655, MAE is 0.6306 * 10-30m3/atom. The R-square of the best model of atomic average energy is 0.9724, MAE is 0.1466 eV / atom. The R-square of the best model of atomic formation energy is 0.9880, MAE is 0.0732 eV / atom. The development of machine learning algorithm in the field of materials can solve the consumption of time and financial resources caused by the huge data calculation of traditional mathematical models. Moreover, with the improvement of the accuracy of the prediction results, the prediction algorithm can guide the experiment to a certain extent in the future and can avoid a lot of redundant and useless work. It will be a trend in the field of materials in the future to predict the properties of materials in advance and dig out the properties of unknown alloys through machine learning.