Multifidelity Machine Learning for Molecular Excitation Energies

Area of Science:

• Computational chemistry
• Quantum chemistry
• Machine learning applications

Background:

• Calculating molecular excited states accurately and quickly is a significant challenge.
• Accurate excitation energies are crucial for understanding energy funnels in molecular aggregates.
• High computational costs for generating accurate training data limit machine learning applications in this field.

Purpose of the Study:

• To develop a cost-effective machine learning approach for predicting molecular vertical excitation energies.
• To demonstrate the efficacy of multifidelity machine learning in achieving high accuracy with reduced computational expense.
• To apply the multifidelity approach to benzene, naphthalene, and anthracene.

Main Methods:

• Utilizing multifidelity machine learning by combining limited high-accuracy data with abundant low-accuracy data.
• Training and testing models on vertical excitation energies to the first excited state.
• Employing conformations from classical molecular dynamics and density functional based tight-binding simulations.
• Comparing the performance of the multifidelity model against a model trained solely on high-cost data.

Main Results:

• The multifidelity machine learning model achieved accuracy comparable to models trained exclusively on high-cost data.
• A computational cost reduction exceeding a factor of 30 was observed in benchmark calculations.
• The approach demonstrated significant potential for further gains with higher accuracy data.

Conclusions:

• Multifidelity machine learning offers a computationally efficient strategy for accurate prediction of molecular excitation energies.
• This method effectively overcomes the data generation cost barrier in applying machine learning to complex chemical systems.
• The successful application to benzene, naphthalene, and anthracene validates its utility for larger molecular systems.