Machine-Learning Coupled Cluster Properties through a Density Tensor Representation.

Area of Science:

• Quantum mechanics
• Computational chemistry
• Machine learning

Background:

• Machine learning (ML) offers potential solutions for computationally intensive problems in electronic structure theory.
• Current ML applications in electronic structure primarily focus on learning molecular energies from geometric data, often requiring extensive training datasets.
• Solving the electronic Schrödinger equation and managing complex wave function parameter spaces remain significant computational challenges.

Purpose of the Study:

• To develop a general, systematically improvable machine learning scheme for predicting arbitrary molecular properties.
• To create a more data-efficient ML model for quantum mechanical calculations.
• To leverage existing ML frameworks with novel representations for enhanced accuracy.

Main Methods:

• Combining the t-amplitude tensor representation with a new reduced density matrix representation.
• Developing a wave function-based machine learning approach inspired by canonical quantum mechanical equations.
• Training and validating the model on small molecules for electronic energy and dipole moment predictions.

Main Results:

• Achieved quantitative accuracy in predicting electronic energy and dipole moments for small molecules.
• Demonstrated the model's effectiveness using only a few dozen training points per system, significantly reducing data requirements.
• Validated a novel combination of ML representations for molecular property prediction.

Conclusions:

• The proposed wave function-based ML scheme offers a data-efficient and accurate method for calculating molecular properties.
• This approach provides a systematically improvable pathway for applying ML to complex quantum mechanical problems.
• The combination of t-amplitude tensors and reduced density matrices represents a significant advancement in ML for computational chemistry.