基于多任务学习的改性双基推进剂的综合性能预测

Accurate prediction of the comprehensive properties for CMDB propellants based on multi-task learning

为满足改性双基推进剂多性能的综合预测需求,本研究提出基于多任务学习的机器学习策略,综合考虑推进剂组分、含量、压强和的粒度对目标性能的影响,首次构建了包含燃速、比冲、特征速度、摩擦感度和撞击感度在内的RDX-CMDB推进剂综合性能预测模型。通过网格寻参模式优化模型,结合十折交叉验证法比较了十种机器学习算法的建模效果。其中,极限梯度提升回归模型预测性能最优,平均R^(2)可达0.9997;在对6个外部样本的测试中,该模型对5个目标性能的预测误差均在5%以内。结果表明,本研究提出的多任务机器学习模型可在试验样本量不足的情况下,实现推进剂的多个目标性能准确预测,对推进剂的综合性能优化和配方设计具有理论指导意义。

In order to meet the demand for comprehensive prediction of multiple properties of composite modified double base(CMDB) propellants,this study firstly proposed a machine learning strategy based on multi-task learning(MTL) to construct a comprehensive prediction model for RDX-CMDB propellants including burning rate,specific impulse,characteristic velocity,friction sensitivity and impact sensitivity,in which the effects of propellant composition,content,pressure and particle size on the target's performance were taken into account from various aspects.A comparison of ten machine learning algorithms using grid-search parameter tuning and 10-fold cross validation showed that the extreme gradient boosting(XGB) regression model performed the best with an average R^(2) of 0.9997,and the model was validated on 6 external samples with the relative prediction error less than 5%,confirming the good reliability and robustness of our model.In conclusion,the MTL model proposed in this work effectively solves the problem of insufficient propellant sample size.It can simultaneously and accurately predict multiple target performances of RDX-CMDB propellants,and is of theoretical significance in guiding the comprehensive performance optimization and formulation design of propellants.