机器学习辅助U-Mo合金等温分解参数设计

Machine Learning Assisted Design of Isothermal Decompositon Parameters of U-Mo Alloy

摘要:

U-Mo合金是一种在研究试验堆中极具发展潜力的弥散燃料。提高氢化-脱氢制粉效率是通过粉末冶金方法高效制备U-Mo合金弥散燃料的前提。优化均匀化处理温度、等温时效温度、等温时效时间、Mo含量等参量有利于增加U-Mo合金α相含量,进而提高U-Mo合金制粉效率。机器学习辅助设计材料的方法能够大幅减少昂贵耗时的试验次数,提高材料研发效率。本文将机器学习方法应用于U-Mo合金等温分解参数的快速设计,以合金硬度为设计指标,基于少量数据建立了合金硬度与上述参数之间的机器学习支持向量机模型。在对硬度预测的基础上,比较了基于预测值和基于预期提高的两类实验设计算法在优化效率方面的差异。结果表明,基于预期提高的实验设计算法通过少量迭代试验能够明显提高硬度,而基于预测值的设计算法对硬度提高不明显。应用上述机器学习辅助设计方法,通过4次实验成功地确定了该合金等温分解最佳参数组合为时效温度为565 °C,时效时间20小时以上,均匀化处理温度为900~950 °C,Mo含量为6wt.%,在该工艺窗口下处理的合金硬度最高,制粉率最高。本研究对利用机器学习方法快速优化U基合金工艺参数进行了初步尝试,这类基于数据的方法能够有效提高材料研发效率。

Abstract:

U-Mo alloy is with great development potential as a kind of dispersive fuel in research and test reactors. Improving the efficiency of powder obtention via hydride-dehydride process is a prerequisite for efficient powder metallurgy preparation of U-Mo alloy dispersion fuels. Optimizing parameters such as homogenization temperature, isothermal aging temperature, isothermal aging time, and Mo content is beneficial to increase the α-phase content of U-Mo alloys, thereby improving the efficiency of the power obtention of U-Mo alloy. Machine learning aided design of materials can greatly reduce the trials of expensive and time-consuming experiments and improve the efficiency of material development. In this paper, a machine learning method is applied to the rapid design of isothermal decomposition parameters of U-Mo alloys. With the hardness of the alloy as a design index, a machine learning support vector machine (SVM) model between the alloy hardness and the above parameters is established based on a small amount of data. Based on the prediction of hardness, the differences in optimization efficiency between the two types of experimental design algorithms based on predicted values and based on expected improvement are compared. The results show that the experimental design algorithm based on the expected improvement can significantly improve the hardness through a small number of iterative experiments, while the design algorithm based on the predicted value does not significantly improve the hardness. Using the above-mentioned machine learning aided design method, the optimal parameter combination for isothermal decomposition of the alloy was successfully determined through 4 experiments. When the aging temperature is 565 °C, the aging time is more than 20 h, the homogenization temperature is 900~950 °C, and the Mo content is 6wt.%, the hardness of the alloy processed is the highest, and the powder obtention rate is the highest. This study made a preliminary attempt to use machine learning methods to quickly optimize U-based alloy process parameters. Such data-based methods can effectively improve the efficiency of material development.