Prediction of Density and Viscosity of Biofuel Compounds Using Machine Learning Methods

In the present work, temperature dependent models for the prediction of densities and dynamic viscosities of pure compounds within the range of possible alternative fuel mixture components are presented. The proposed models have been derived using machine learning methods including Artificial Neural Networks and Support Vector Machines. Experimental data used to train and validate the models was extracted from the DIPPR database. A comparison between models using an ample range of molecular descriptors and models using only functional group count descriptors as inputs was performed, and consensus models were created by testing different combinations of the individual models. The resulting consensus models’ predictions were in agreement with the available experimental data. Comparisons were also made between predictions of our models and correlations validated by the DIPPR staff. Our models were used to predict densities and dynamic viscosities of compounds for which no experimental data exists. Our models were also used to estimate other properties such as kinematic viscosities, critical temperatures, and critical pressures for compounds in the database. Finally, predictions were used to study the main trends of density and viscosity at the aforementioned temperatures as a function of the number of carbon atoms for chemical families of interest.本工作提出了可用于预测潜在替代燃料混合组分范围内纯化合物密度与动力粘度的温度依赖型模型。所提出的模型基于机器学习方法构建，涵盖人工神经网络（Artificial Neural Networks）与支持向量机（Support Vector Machines）两类算法。用于训练与验证模型的实验数据均提取自DIPPR数据库。本研究开展了两类模型的对比：一类以全面的分子描述符作为输入，另一类仅以官能团计数描述符作为输入；随后通过测试单一模型的不同组合构建了共识模型。所得到的共识模型预测结果与已公开的实验数据吻合良好。本研究还将所提模型的预测结果与经DIPPR团队验证的物性关联式进行了对比。利用所提模型，可对尚无实验数据的化合物的密度与动力粘度进行预测。此外，本研究还利用所提模型估算了数据库中化合物的其他物性，包括运动粘度、临界温度与临界压力。最后，基于预测结果，本研究分析了目标同系物在上述温度下，密度与粘度随碳原子数变化的主要趋势。