Mimyria: Machine-Learned Vibrational Spectroscopy for Aqueous Systems Made Simple

Area of Science:

• Computational Chemistry
• Spectroscopy
• Machine Learning

Background:

• Vibrational spectroscopy links molecular dynamics (MD) simulations and experiments.
• Routine use in condensed-phase systems is limited.
• Need for automated, efficient workflows.

Purpose of the Study:

• Introduce mimyria, a unified workflow for vibrational spectroscopy.
• Develop and validate atom-resolved machine-learning targets (APT for IR, PGT for Raman).
• Enable data-efficient and reliable spectral generation from MD.

Main Methods:

• Orchestrate electronic-structure calculations and train atom-resolved ML response models.
• Introduce and compute the polarizability gradient tensor (PGT) for Raman spectroscopy.
• Validate ML models against ab initio calculations and spectral accuracy.

Main Results:

• IR and Raman spectra converge rapidly with small training sets.
• Spectral agreement improves faster than root-mean-square error (RMSE).
• Established guidelines for achieving spectral fidelity and early stopping criteria.

Conclusions:

• mimymria enables data-efficient, quantitatively reliable vibrational spectroscopy.
• Connects model-level errors to observable-level accuracy.
• Automated framework integrates response-tensor learning and spectral validation.