Estimating PC-SAFT Pure-Component Parameters by Using Graph Neural Networks: Application t

A dataset of 1842 molecules used to train a Graph Neural Network model for predicting PC-SAFT equation-of-state parameters from molecular structure. The model was developed by Wildson Bernardino de Brito Lima and last updated in May 2026, achieving a liquid density mean absolute percentage error of 8.21% on a test set of 1642 unseen molecules. It demonstrated improved accuracy over traditional group contribution methods for complex solvents like ionic liquids and deep eutectic solvents.

Use Cases

• Predict PC-SAFT pure-component parameters for novel molecules based on molecular structure.
• Estimate thermodynamic properties like liquid density and vapor pressure for ionic liquids.
• Parametrize deep eutectic solvents for process design without relying on experimental data.
• Benchmark new machine learning models against the reported GNN performance metrics.

Strengths

• Model trained on 1842 molecules and validated on a test set of 1642 unseen molecules.
• Demonstrated superior performance over a traditional method, reducing density MAPE from 7.17% to 3.79% on a subset of 122 molecules.
• Achieved specific, low error rates for complex solvents, e.g., a density MAPE of 0.20% for the ionic liquid [Bmim][Tf2N].

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect bias inherent to the source training molecules from the ThermoML Archive.

Provenance

Source

figshare

Collection Method

Parameters derived via a Graph Neural Network model trained on molecular structures.

Freshness

Last updated 2026-05-28 06:14:07