Trained machine-learning models for molecular absorption-spectrum prediction

Abstract

This record contains trained machine-learning model weights developed for predicting molecular absorption spectra and associated spectrum-normalization constants from molecular structure and ground-state electron density.

Four trained models are provided:

cnn_model_spectra.pt — convolutional neural network model for predicting molecular absorption spectra from three-dimensional ground-state electron-density representations.

cnn_model_normconst.pt — convolutional neural network model for predicting the normalization constant used for molecular absorption spectra from three-dimensional ground-state electron-density representations.

mace_model_spectra.pth — MACE model for predicting molecular absorption spectra directly from molecular atomic structure.

schnet_model_spectra.pt — SchNet model for predicting molecular absorption spectra directly from molecular atomic structure.

The models were developed using molecules from the QM7 molecular dataset together with electronic-structure calculations used to generate ground-state electron densities and absorption spectra. The spectrum-prediction models were trained to reproduce the calculated spectral response over the energy/frequency representation used in the associated study.

These files contain trained model parameters intended for inference and reproducibility of the associated machine-learning results. Model architectures, preprocessing procedures, inference scripts, data splits, normalization procedures and examples of model use are provided separately in the associated software repository/publication.

The CNN density-to-spectrum model provides a surrogate mapping from ground-state electron density to molecular spectral response, while the SchNet and MACE models provide structure-to-spectrum models using atomic coordinates and species as inputs. The normalization-constant CNN predicts the scaling quantity required to recover consistently normalized spectra.

These model weights are released to support reproducibility, reuse and comparison of machine-learning approaches for molecular spectroscopy.