Mimyria [J. Chem. Theory Comput., 2026], is a modular framework for predictive vibrational spectroscopy from molecular dynamics simulations. It combines electronic-structure calculations, machine-learning response models, and automated spectral analysis within a unified workflow for infrared (IR) and Raman spectroscopy in condensed phases.

The framework enables the efficient calculation of fully anharmonic, temperature-dependent vibrational spectra from large-scale molecular dynamics simulations while retaining electronic-structure accuracy. To this end, mimyria implements machine-learning models for atom-resolved spectroscopic response properties, including atomic polar tensors (APTs) for IR spectroscopy and polarizability gradient tensors (PGTs) for Raman spectroscopy.

Starting from an existing machine-learning potential and molecular dynamics trajectory, mimyria provides the complete pipeline required to generate predictive vibrational spectra. The software automates electronic-structure reference calculations, response-tensor generation, machine-learning training, validation, and the final computation of IR and Raman spectra from MLMD simulations. By validating models directly at the spectral level and connecting machine-learning errors to observable spectral accuracy, mimyria enables data-efficient and quantitatively reliable spectroscopy simulations for complex molecular systems.

The repository contains training workflows, spectral-analysis tools, and example applications for IR and Raman spectroscopy from machine-learning molecular dynamics simulations. Feel free to explore!

Github

The Atomic Polar Tensor Neural Network (APTNN) has been introduced to compute IR spectra from molecular dynamics simulations leveraging machine learning [J. Chem. Theory Comput., 2023]. It has also been used include electric fields in machine learning molecular dynamics simulations in a perturbative approach [Nat. Commun., 2024].

The APTNN is now considered legacy software, as its functionality has largely been integrated and extended within the Mimyria framework, which provides a more general and automated workflow for predictive vibrational spectroscopy.

The repository nevertheless remains available and contains scripts and examples for training APTNN models and conducting electric-field simulations. Feel free to explore!

Github