Predicting Thermodynamic Properties of Alkanes by High-Throughput Force Field Simulation and Machine Learning
Predicting Thermodynamic Properties of Alkanes by High-Throughput Force Field Simulation and Machine Learning
复制标题
通过高通量力场模拟和机器学习预测烷烃的热力学性质
DOI：
10.1021/acs.jcim.8b00407
复制
发表时间：
2018-12-01
影响因子：
5.6
通讯作者：
Sun，Huai
中科院分区：
文献类型：
--
作者：
Gong，Zheng;Wu，Yanze;Sun，Huai
分子热力学性质的知识对于化学工艺设计和新材料的开发是必不可少的。实验测量通常昂贵且不环保。在过去，由于缺乏一致的力场和模拟协议，使用分子模拟的研究集中在特定类别的分子上。为了解决这个问题，我们已经开发了一个高通量的力场模拟（HT-FFS）程序相结合，最近开发的一般力场与验证的模拟协议来计算大量分子的热力学性质。这个程序适用于计算液体密度，汽化热，热容，汽液平衡曲线，临界温度，临界密度和表面张力的烷烃范围很广。的预测与现有的实验数据在准确性和精度方面吻合良好，表明HT-FFS是一种有效的方法来补充实验测量。此外，HT-FFS生成的大量数据为机器学习奠定了基础。我们已经开发了一个人工神经网络，它证明了使用机器学习模型将预测扩展到模拟之外的可行性。
Knowledge of the thermodynamic properties of molecules is essential for chemical process design and the development of new materials. Experimental measurements are often expensive and not environmentally friendly. In the past, studies using molecular simulations have focused on a specific class of molecules, owing to the lack of a consistent force field and simulation protocol. To solve this problem, we have developed a high-throughput force field simulation (HT-FFS) procedure by combining a recently developed general force field with a validated simulation protocol to calculate thermodynamic properties for large number of molecules. This procedure is applied to calculate liquid densities, heats of vaporization, heat capacities, vapor-liquid equilibrium curves, critical temperatures, critical densities and surface tensions for a wide range of alkanes. The predictions agree well with available experimental data in terms of accuracy and precision, demonstrating that HT-FFS is a valid approach to supplementing experimental measurements. Furthermore, the large amount of data generated by HT-FFS lays a foundation for machine learning. We have developed an artificial neural network that demonstrates the feasibility of expanding predictions beyond simulation using a machine learning model.