Elsevier

Computational Materials Science

Volume 155, December 2018, Pages 48-54
Computational Materials Science

Machine leaning aided study of sintered density in Cu-Al alloy

https://doi.org/10.1016/j.commatsci.2018.07.049Get rights and content

Abstract

The mechanical properties of powder metallurgy (PM) materials are closely related to their density. In this case we demonstrate an approach of utilizing machine-learning algorithms trained on experimental data to predict the sintered density of PM materials. Descriptors were selected from the features including processing parameters, chemical composition, property of raw materials and so on. And the training data are collected by the experimental setup in our lab and the literatures on five kinds of P/M alloys. The multilayer perceptron model (MLP) outperformed other four regression and neutral network models with high coefficient of correlation and low error. The sintered density predicted by MLP model agreed well with the experimental data with a tolerable error less than 0.028, which confirms its capability over P/M materials design procedures. Then the obtained MLP model is used for Cu-9Al P/M alloy to guide selecting the processing parameters to reach the expected sintered density of 0.88. The Cu-9Al powders were fabricated with the predicted parameters including the specific shape factor, particle size, pressing pressure and sintering temperature, and the obtained relative sintered density is 0.885.

Introduction

Powder metallurgy (PM), with the descriptors of energy-efficient near-net forming and no (less) pollution, plays an irreplaceable role in the materials manufacturing [1], [2]. The sintered density is one of the key factors that determine the mechanical properties of powder metallurgy parts [3]. An increase of the porosity from 0 to 5% decreases the tensile strength by 35% [4]. Ductility and resistance to fatigue show higher sensitivity to porosity [5]. At present, most of the sintered density prediction depends on the priori knowledge and trial and error experiments, which greatly decelerate the efficiency of new materials discovery.

Therefore, a more efficient method is urgently needed to predict the sintered density. The Materials Genome Initiative (MGI) [6] and Integrated Computational Materials Engineering [7] (ICME) are deemed to accelerating the materials design process. Xue et al. [8] discovered a new NiTi-based shape memory alloy with targeted transformation temperatures by Support Vector Regression method. Raccuglia [9] used support vector machine methods to predict the viability of untested reactions with experimental failed data collected in the lab and the accuracy of prediction is nearly 90%.

Machine learning methods have been used to predict the sintered density in recent years. Jabar et al. [10] used artificial neural network (ANN) to predict the sintered density of BaTiO3 and the mean error is less than 0.06. Varol et al. [11] used ANN to predict the effect of reinforcement on density, hardness and tensile strength of B4Cp/Al2024 composites, with the mean absolute percentage error (MAPE) obtained less than 2.15%. Canakci et al. [12] used ANN to predict green density, sintered density and hardness of Al2O3/Al metal matrix composites (MMCs) successfully with MAPE less than 5.53%. Azadbeh et al. [13] used response surface methodology models to predict the density, electrical resistivity and hardness of Cr–Mo prealloyed sintered steels, the predicted values agreed well with the experimental ones. The previous investigations, however, focused studies solely on single kind of material for prediction without taking into account of the raw material and the phase transformation, and didn’t compare the prediction accuracy of different machine learning methods.

In the present work, machine learning algorithms were used to generate the models to estimate the sintered density of metal compacts by Powder metallurgy. And the obtained model is applied here to predict the sintered density Cu-9Al alloy as well as provide guidance for selection of processing parameters to reach the expected target.

Section snippets

Procedure of modeling

The feedback mechanism of the machine learning modeling and sintering optimization process are described in Fig. 1. Machine-learning models generated from experimental data are used to guide selection of processing parameters to reach the expected sintered density.

Model construction

A comparison of the models training results with the experimental ones is shown in Fig. 3, where the x-axis is the relative sintered density from experiments and the y-axis from models prediction. For an ideal model, the data points are more close to diagonal line. In this case MLP model performs better than the others.

In this case, the coefficient of correlation (R), mean absolute error (MAE) and root mean squared error (RMSE) [35], [36] are used to evaluate the feasibility and validation of

Machine leaning model guides the selection of parameters

The obtained MLP model is applied here to predict the sintered density Cu-9Al alloy as well as provide guidance for selection of processing parameters to reach the expected target without the time consuming trial-and-error method. In this case, we focus solely on the data of Cu-Al alloys in MLP predicted values for investigating the correlations between each descriptor and the sintered density, with chemical composition and sintering atmosphere negleted. The results are shown in Fig. 6. The

Conclusions

In this work, five algorithms are used for modeling. By comparing five prediction models, the fitting of the MLP model was the best with the predicted value of MLP model was in well agreement with the experimental value, and the maximum error value does not exceed 0.028. The MLP model is applied to predict the sintered density of Cu-Al alloy as well as provide guidance for selection of processing parameters to reach the expected target. The selected Cu and Al powders for Cu-9Al alloy were die

CRediT authorship contribution statement

ZHD and HQY: conceived and designed the study. ZHD and HQY wrote the manuscript. ZHD, KQZ, TZ, BX and QJZ collected data. ZHD and XJ performed descriptor selection and model construction. CZ calculated volume change induced by the phase transformation. ZHD, KQZ, TZ performed the experiments. ZHD, HQY and XHQ reviewed and edited the manuscript. All authors read and approved the manuscript.

Acknowledgement

The National Key Research and Development Program (2016YFB0700503), the National High Technology Research and Development Program of China (2015AA03420), Beijing Science and Technology Plan (D16110300240000), National Natural Science Foundation of China (No. 51172018) and Kennametal Inc. are greatly acknowledged.

References (37)

  • H.L. Hao et al.

    The effect of trace amount of Ti and W on the powder metallurgy process of Cu

    J. Alloys Comp.

    (2016)
  • P.J. Denny

    Compaction equations: a comparison of the Heckel and Kawakita equations

    Powder Technol.

    (2002)
  • R. Meenal et al.

    Selvakumar. Assessment of SVM, Empirical and ANN based solar radiation prediction models with most influencing input parameters

    Renew. Energy

    (2018)
  • X. Jiang et al.

    An materials informatics approach to Ni-based single crystal superalloys lattice misfit prediction

    Comput. Mater. Sci.

    (2018)
  • D. Caballero et al.

    Prediction of pork quality parameters by applying fractals and data mining on MRI

    Food Res. Int.

    (2017)
  • M.W. Gardner et al.

    Artificial neural networks (the multilayer perceptron)—a review of applications in the atmospheric sciences

    Atmos. Environ.

    (1998)
  • Y.Y. Li et al.

    Progress in high density powder metallurgy forming technology

    Mater. Sci. Eng. Powder Metall.

    (2005)
  • P.Y. Huang

    The principle of Powder Metallurgy

    (1997)
  • Cited by (0)

    View full text