Trained machine-learning models for molecular absorption-spectrum prediction

Abstract

This record contains five trained machine-learning models for predicting molecular absorption spectra and associated spectrum-normalization constants from molecular atomic structure and three-dimensional ground-state electron density.

The following model files are provided:

cnn_model_spectra.pt — a three-dimensional convolutional neural network (CNN) for predicting normalized molecular absorption spectra from ground-state electron-density grids.

cnn_model_normconst.pt — a three-dimensional CNN for predicting the spectrum-normalization constant from ground-state electron-density grids. Its matching training configuration is required to convert model outputs to the original spectrum-sum units.

mace_model_spectra.pth — trained MACE model parameters for predicting normalized molecular absorption spectra from atomic species and Cartesian coordinates.

schnet_model_spectra.pt — a SchNet checkpoint bundle for predicting normalized molecular absorption spectra from atomic species and Cartesian coordinates.

dimenetpp_model_spectra.pt — a DimeNet++ checkpoint bundle for predicting normalized molecular absorption spectra from atomic species and Cartesian coordinates.

The SchNet and DimeNet++ files contain trained weights, model identifiers, architecture settings and training configuration. These bundles allow the architectures to be reconstructed using the associated software; they require the repository code and its dependencies for inference.

The models were developed using molecules from the QM7 dataset (http://quantum-machine.org/datasets/) and electronic-structure calculations of ground-state electron densities and absorption spectra. Spectrum predictions are evaluated over a 900-bin energy window from 0.0 to 0.4495 atomic units, with a spacing of 0.0005 atomic units, and are normalized to unit sum within this window. The normalization CNN predicts the sum of the corresponding 900 reference-spectrum values, enabling recovery of the calculated spectrum's intensity scale when used with a consistently normalized spectral prediction.

The density-to-spectrum CNN learns a surrogate mapping from ground-state electron density to molecular spectral response. SchNet, DimeNet++ and MACE learn structure-to-spectrum mappings from atomic species and coordinates. These models support comparison of density-based and structure-based representations for molecular spectroscopy.

Model definitions, preprocessing procedures, inference scripts, training and data-split documentation, normalization conventions and examples are available in the associated software repository: https://github.com/stfc-ai4s/QM7-Absorption-ML

These trained models are released to support inference, reproducibility, reuse and comparison of machine-learning approaches for molecular spectroscopy within the scope of the associated study.