Computing Bulk Phase IR Spectra from Finite Cluster Data via Equivariant Neural Networks

Area of Science:

• Computational chemistry
• Spectroscopy
• Machine learning

Background:

• Accurate infrared (IR) spectra calculation from molecular dynamics (MD) simulations is vital for structural dynamics and simulation benchmarking.
• Machine learning (ML) has accelerated these calculations, but using finite-cluster data for condensed-phase IR spectra is an open area.
• Bridging the gap between finite-cluster data and bulk properties in IR spectral calculations is essential.

Purpose of the Study:

• To investigate if ML models trained solely on electronic structure calculations of finite-size clusters can reproduce bulk IR spectra.
• To establish a connection between finite-cluster data and macroscopic IR spectral properties.
• To demonstrate the feasibility of predicting condensed-phase IR spectra using limited, localized data.

Main Methods:

• Utilized electronic structure calculations on finite-size molecular clusters.
• Employed an equivariant neural network architecture.
• Targeted the atomic polar tensor as the key training property for the ML model.
• Validated the model's performance against known bulk IR spectra of liquid water.

Main Results:

• The equivariant neural network accurately reproduced the bulk IR spectrum of liquid water.
• Demonstrated that finite-cluster electronic structure data is sufficient for predicting bulk IR spectra.
• Established the atomic polar tensor as a reliable property for training ML models for IR spectra prediction.

Conclusions:

• Machine learning models trained on finite-cluster data can successfully predict bulk IR spectra.
• This approach provides a computationally efficient method for obtaining condensed-phase IR spectra.
• The study validates the use of atomic polar tensors and equivariant neural networks for this task, linking microscopic calculations to bulk phenomena.