文章信息/Info

Title:

Machine Learning Prediction Model for Microstructure-Tensile Properties Relationship of Superalloys

文章編號:

1674-3962(2022)11-0938-09

作者:

劉芳寧; 王越; 孫瑞俠

（中國航發北京航空材料研究院，北京 100095）

Author(s):

LIU Fangning;  WANG Yue;  SUN Ruixia

(AECC Beijing Institute of Aeronautical Materials, Beijing 100095, China)

關鍵詞:

機器學習; 相含量; 力學性能; 高溫合金; 材料科學

Keywords:

machine learning;  phase content;  mechanical properties;  superalloy;  material science

分類號:

TP181；TG132.3+3

DOI:

10.7502/j.issn.1674-3962.202101024

文獻標志碼:

A

摘要:

采用傳統“試錯法”揭示高溫合金微觀組織和性能之間的關系具有成本高、周期長等特點，難以滿足材料設計、研發和應用的快速發展。以高溫合金K4169為基礎，采用機器學習的方法建立了材料微觀組織和力學性能之間關系的模型。首先設計實驗，通過改變合金成分和熱處理制度，獲取了70組微觀組織變化數據；其次通過室溫和高溫拉伸實驗對不同微觀組織的合金力學性能進行測量，分析了成分和熱處理制度變化對合金室溫、高溫力學性能的影響；最后分別采用支持向量回歸、隨機森林回歸、K�沧罱�鄰回歸、多層感知器4種算法建立預測模型，預測了微觀組織中γ相、γ′相、γ″相、δ相、Laves相和碳化物含量對合金室溫、高溫拉伸性能的影響，并采用交叉驗證的方式驗證了模型的準確性。結果表明，多層感知器模型對合金室溫拉伸強度的預測結果均方誤差為0.17、平均絕對誤差為0.32、相關系數為0.95、決定系數為0.85，對合金高溫拉伸強度的預測結果均方誤差為0.14、平均絕對誤差為0.29、相關系數為0.97、決定系數為0.91，與其余3種算法建立的模型相比，多層感知器模型的預測結果更準確。

Abstract:

The traditional “trial-and-error” method is used to reveal the relationship between microstructure and properties of superalloys, which is difficult to meet the rapid development of material design, research and development and application due to its high cost and long cycle. In this paper, a model for the relationship between microstructure and properties of superalloy K4169 was established by machine learning. Firstly, 70 groups of microstructures were obtained by changing the alloy composition and heat treatment system. Secondly, the mechanical properties of the alloys were measured by tensile experiments at room temperature and high temperature, and the effects of composition and heat treatment on the mechanical properties were analyzed. Finally, support vector machine regression (SVR), random forest regression (RFR), K-nearest neighbor node regression (KNR) and multi-layer perceptron (MLP) were used to establish prediction models to predict the effects of γ phase, γ′ phase, γ″ phase, δ phase, Laves phase and carbide content on tensile properties at room and high temperature. The accuracy of the model was verified by cross validation. The results show that the mean squared error of the MLP model for the prediction result of the alloy room temperature tensile strength is 0.17, the mean absolute error is 0.32, the correlation coefficient is 0.95, and the decision coefficient is 0.85.And the mean squared error for the alloy high-temperature tensile strength is 0.14, the mean absolute error is 0.29, the correlation coefficient is 0.97, and the decision coefficient is 0.91.Compared with the other three models, the prediction results of MLP are more accurate.