基于Bayesian采样主动机器学习模型的6061铝合金成分精细优化

Composition Refinement of 6061 Aluminum Alloy Using Active Machine Learning Model Based on Bayesian Optimization Sampling

结合高通量材料制备实验与基于Bayesian优化采样策略的主动学习方法,开发了有效的机器学习模型来描述合金元素组成与硬度之间的关系,并分析关键微量元素含量对硬度的影响。研究发现,经过3轮迭代64个铝合金样品建模后,Bayesian取样策略方法的预测硬度误差为4.49 HV(7.23%),远低于应用人工经验采样法的机器学习模型误差9.73 HV(15.68%),且当铝合金中的Mg和Si比值Mg/Si在1.37~1.72时,具有较高的合金硬度。通过在6061铝合金标准名义成分范围内进行成分精细优化以及性能调控,为工业上提高产品质量提供了可实现的策略.

Within the China standard(6061 GB/T 3190-2008)of the aluminum alloy 6061,there are a wide range of alloy compositions having multiple trace elements.From the viewpoint of scientific research and quality control in industries,it is important to understand the relationship between the different potential compositions and corresponding mechanical properties of the aluminum alloy 6061.In this work,high-throughput experiments on materials synthesis and an active-learning framework based on the Bayesian optimization sampling process were combined to develop effective machine learning(ML)models to describe the relationship between the composition and hardness of aluminum 6061 alloys.In this work,>100 alloys with ML designed compositions were synthesized and their hardness data were obtained through high-throughput experiments.The composite ML features were introduced by combining elementary material properties and chemical compositions of alloys and were selected subsequently according to their importance and correlation among features.The efficiencies of two sampling strategies were compared in guiding the iterative experiments:manual sampling based on empirical experience and Bayesian optimization sampling trained within the active-learning framework using the efficient global optimization and knowledge gradient algorithms.These ML models were updated iteratively until the prediction accuracy approached the experimental error.Specifically,the error in the hardness values predicted by the Bayesian model using 64 aluminum alloy samples after three rounds of iterations was 4.49 HV(7.23%),which is much lower than the error predicted by the empirical sampling method(9.73 HV;15.68%).The results show that Bayesian optimization sampling accelerates the optimization of alloys property more efficiently than manual empirical sampling.Finally,the machine learning models using Bayesian sampling were interpreted using the Shapley additive explanations method and analysis of the partial dependence plot discuss the effects of various trace alloying elements and composite ML features on the hardness of the aluminum alloys.It was found that the hardness value of the aluminum alloys became large when the ratio between Mg and Si(Mg/Si)was between 1.37 and 1.72.In addition,the machine learning models suggested that the lattice distortion,cohesive energy,configurational entropy,and shear modulus were positively proportional to the hardness of the alloy.This work demonstrated that active-learning-guided high-throughput experiments on composition refinement can not only improve the performance and quality control of aluminum 6061 alloys within its standard nominal composition range as used in industry but also provide a feasible approach for the design and property optimization of other multialloy materials.