Machine Learning Densities, Detonation Velocities, and Formation Enthalpies of Energetic Materials Using Quantum Chemistry Descriptors

The prediction of detonation parameters is a challenging task requiring to bridge the gap between microscopic molecular features and macroscopic materials properties. Whereas traditional routes are concerned with empirical equations, we present a machine learning approach to this task here. Our approach capitalizes on molecular descriptors from high-level quantum chemistry as input to produce models fitted against experimental reference data to model three key quantities: the crystalline density, the detonation velocity, and the heat of formation. To determine the detonation products, we use a new nonempirical product optimization scheme, maximizing the heat release, which is extensible to any molecular composition. We find, for all three properties, that the machine-learned results significantly surpass standard rule-based schemes. Finally, we present an all in silico scheme for predicting detonation velocities, highlighting that this is almost as good as when experimental densities are used as input. In summary, we believe that this work is a major step toward the goal of accurately predicting detonation parameters by showing how to leverage the power of quantum chemistry for this task.爆轰参数的预测是一项极具挑战性的任务，需要弥合微观分子特征与宏观材料性能之间的鸿沟。传统研究方法多依赖经验公式，本文提出了一种机器学习方法来解决该任务。本方法以高精度量子化学得到的分子描述符作为输入，基于实验参考数据拟合得到预测模型，用于对三项关键参数进行建模：晶体密度、爆轰速度以及生成焓。为确定爆轰产物，我们采用了一种全新的非经验产物优化方案，通过最大化释热量实现该目标，该方案可扩展至任意分子组成体系。我们发现，针对上述三项参数，机器学习预测结果均显著优于标准的基于规则的预测方法。最后，我们提出了一种全计算流程用于爆轰速度预测，结果表明该方法的预测精度几乎与以实验密度作为输入时的效果相当。综上，本研究通过展示如何利用量子化学的技术优势解决该任务，我们认为这项工作朝着精准预测爆轰参数的目标迈出了重要一步。