Selected machine learning of HOMO-LUMO gaps with improved data-efficiency.

Area of Science:

• Computational chemistry
• Quantum machine learning
• Materials science

Background:

• Quantum machine learning (QML) models are crucial for predicting molecular electronic properties in organic electronics.
• Current QML models often require large training datasets, limiting their data-efficiency.
• Accurate prediction of properties like HOMO-LUMO gaps is essential for designing new organic electronic materials.

Purpose of the Study:

• To enhance the data-efficiency of QML models for molecular electronic properties.
• To investigate the impact of partitioning training data into chemical classes on QML model performance.
• To reduce the computational burden for screening large molecular libraries.

Main Methods:

• Partitioning organic molecules from QM7 and QM9 datasets into three distinct chemical classes.
• Training independent QML models on each chemical class.
• Evaluating QML models based on mean absolute prediction errors for band-gaps at GW and hybrid DFT levels.
• Comparing the data-efficiency of class-selected QML models against randomly trained QML models and Δ-QML models.

Main Results:

• Class-specific QML models achieved mean absolute prediction errors of approximately 0.1 eV.
• These models required up to an order of magnitude fewer training molecules compared to models trained on random data.
• Class-selected QML models demonstrated superior data-efficiency compared to Δ-QML models.
• The identified chemical classes were: aromatic rings/carbonyl groups, single unsaturated bonds, and saturated bonds.

Conclusions:

• Partitioning training data into relevant chemical classes before QML model training significantly improves data efficiency.
• This strategy reduces the number of molecules required for accurate predictions, lowering computational costs.
• Selected QML offers a promising approach for high-fidelity, computationally efficient quantum property screening of large molecular libraries.