Potential Energy Surface Data for Water and Phenol Clusters via Graph Neural Networks

A 187.3 MB dataset supporting a transferable framework for predicting potential energy surfaces in hierarchically structured chemical systems. Siqi Chen developed the FB-GNN-MBE model, which integrates fragment-based graph neural networks with many-body expansion theory, and the data was last updated in April 2026. It includes benchmarks for water, phenol, and mixture systems, demonstrating chemical accuracy for two-body and three-body energy predictions.

Use Cases

• Train graph neural networks to predict two-body (2B) and three-body (3B) interaction energies from fragment-based molecular graphs.
• Fine-tune pre-trained GNN models (student networks) on uniform-density water cluster data for transfer learning across system sizes.
• Benchmark the accuracy of energy predictions against first-principles quantum mechanical models for water and phenol dimers.
• Apply the many-body expansion (MBE) framework to decompose and predict total system energy from individual fragment energies and interactions.

Strengths

• Dataset size of 187.3 MB indicates substantial computational data for model training and validation.
• Validated for chemical accuracy on benchmark systems including water, phenol, and their mixtures.
• Includes a demonstrated transfer learning protocol between mixed-density and uniform-density water cluster ensembles.

Limitations

• Specific row count and number of unique molecular configurations are unknown.
• Data scope is limited to the studied systems (water, phenol, mixtures) and may not generalize to other molecules without adaptation.
• Relies on the accuracy of the underlying first-principles QM models used for one-fragment energy evaluation.

Provenance

Source

Siqi Chen via figshare.

Collection Method

Generated using a fragment-based graph neural network (FB-GNN) integrated with many-body expansion (MBE) theory, trained on quantum mechanical models.

Freshness

Last updated in April 2026.