Empirical correlations derived from existing experimental data have always played an important role in thermophysical property estimation. These empirical correlations were traditionally developed on the basis of some reliable but often very limited data compilations. Currently, large comprehensive experimental data collections have not only become more readily available, but are also more accessible to systematic studies due to rapid advances in database technologies. On the other hand, machine learning methods capable of recognizing data patterns and establishing quantitative relationships have also advanced significantly.

The NIST Thermodynamics Research Center maintains one of the largest electronic collections of experimental thermophysical data in the world. It currently contains over 4.2 million data points along with the associated experimental uncertainties. Direct access to this information opens unique possibilities for the development of next-generation correlations and prediction methods. The program will build on our existing efforts using Quantitative Structure-Property Relationship (QSPR) methodologies and modern machine learning methods (support vector machines, symbolic regression). Various aspects of data processing such as detection of erroneous data (outliers) and data balancing will also be covered.

 

key words

Thermophysical properties; Correlation; Prediction methods; QSPR; Symbolic regression; Machine learning; Databases; Outlier detection; SVM;

Citizenship:  Open to U.S. citizens

Level:  Open to Postdoctoral applicants