Thermochemical Data Fusion Using Graph Representation Learning.

Area of Science:

• Computational Chemistry
• Materials Science
• Catalysis

Background:

• Big Data applications in catalysis and materials science require large thermochemical databases.
• Existing methods for data fusion involve correcting low-fidelity data to higher accuracy with significant computational cost.
• Mapping thermochemical quantities calculated at various density functional theory (DFT) levels to a uniform, high level of theory is crucial for data fusion.

Purpose of the Study:

• To propose a novel graph theoretical, statistical framework for thermochemical data fusion.
• To enable accurate prediction of thermochemical quantities at high levels of theory using minimal computational resources.
• To develop an explainable and generalizable method for enhancing the accuracy of computational chemistry databases.

Main Methods:

• A graph theoretical, statistical framework was developed for data fusion tasks.
• Subgraph frequencies were utilized as a natural representation for learning data fusion maps.
• Linear models with automated descriptor selection were employed to learn these fusion maps.

Main Results:

• The proposed framework can predict multiple thermochemical quantities at a higher level of theory with an accuracy of 1 kcal/mol.
• Models were trained using a small subset (∼1%) of the QM9 database (133,885 molecules).
• The method demonstrated explainability, generalizability, and included a diagnostic tool for outlier identification.

Conclusions:

• The developed graph-based data fusion method significantly enhances the efficiency and accuracy of creating large thermochemical databases.
• This approach is highly effective for Big Data applications in catalysis and materials science, requiring substantially less data and computation.
• The framework offers a robust and interpretable solution for data fusion in computational chemistry, paving the way for more extensive and reliable databases.