Using Autonomous Outlier Detection Methods for Thermophysical Property Data

The reliability and accuracy of thermophysical
property data are
of central importance for the development of models that describe
these properties. In this work, we compare different autonomous algorithms
for identifying the outliers in an existing database. Therefore, the
comprehensive database on thermophysical property data for the Lennard-Jones
fluid [J. Chem. Inf. Model. 2019, 59, 4248–4265] is used. We focus on homogeneous state
property data at given temperature and density for the pressure p, thermal expansion coefficient α, isothermal compressibility
β, thermal pressure coefficient γ, internal energy u, isochoric heat capacity cv, isobaric heat capacity cp, Grüneisen
coefficient Γ, Joule–Thomson coefficient μJT, speed of sound w, chemical potential μ,
(reduced) Helmholtz energy ã = a/T, and its derivatives ãnm. A comprehensive
comparison of 19 outlier
detection methods is carried out, which provides insights into the
applicability of generic outlier detection algorithms for thermophysical
property data. Different classes of outlier detection algorithms are
included in the study, namely, machine learning, distance-based, density-based,
statistical, ensemble, and model-informed. Two approaches are used
for the method evaluation: in approach (a), the original database
(comprising real outliers) is used. In approach (b), synthetic outliers
are introduced. The results and findings from both approaches are
consistent. Machine learning methods yield in some cases better performance
compared to that of the distance-based, density-based, ensemble, and
statistical methods. The best performance is obtained from the model-informed
method (called MoDOD). The results also provide insights into the
nature of the outliers in the Lennard-Jones database.