Machine-Learning Ice Spectra: From 1 to 256 Features.

Area of Science:

• Computational chemistry
• Materials science
• Spectroscopy

Background:

• Predicting spectroscopic properties of materials like ice is crucial for understanding their behavior.
• Machine learning and structural fingerprints offer promising avenues for accelerating these predictions.

Purpose of the Study:

• To evaluate the efficacy of machine learning models and structural fingerprints in predicting ice's OH vibrational frequencies and 1H chemical shifts.
• To compare the performance of different machine learning models and descriptors.

Main Methods:

• Utilized a large theoretical dataset of 55 ice polymorphs with 1010 DFT data points.
• Employed machine learning models including Message Passing Atomic Cluster Expansion (MACE), ACSF, and SOAP.
• Assessed performance using root-mean-square deviation (RMSD) for chemical shifts and vibrational frequencies.

Main Results:

• The MACE model demonstrated superior performance, achieving an RMSD of 0.06 ppm for chemical shifts and ~10 cm⁻¹ for vibrational frequencies.
• Simpler descriptors like ACSF and SOAP, when combined with appropriate regressors, approached MACE's accuracy.
• A basic H-bond distance descriptor resulted in significantly larger RMSD values compared to MACE.

Conclusions:

• Machine learning, particularly MACE, provides highly accurate predictions for ice's spectroscopic properties.
• While simpler descriptors are less accurate, they offer transparency and may be suitable for specific applications.
• The study highlights the potential of computational methods for advancing ice science.