Reversible Deep Learning for Spectral Analysis

A new study introduces a reversible deep learning model designed for 13C nuclear magnetic resonance (NMR) analysis in chemoinformatics. The model uses a single conditional invertible neural network to establish a bidirectional relationship between molecular structures and their corresponding spectra.

The network is built using i-RevNet style bijective blocks, a feature that allows both the forward map and its inverse to be available by construction. This approach allows the model to be trained to predict a 128-bit binned spectrum code from a graph-based structure encoding. The remaining latent dimensions capture residual variability.

Generation of Candidate Structures

During inference, the same trained network is inverted to generate candidate structures from a spectrum code. This process explicitly represents the one-to-many nature of spectrum-to-structure inference. The results show that the model is numerically invertible on trained examples, achieves spectrum code prediction above chance, and produces coarse but meaningful structural signals when inverted on validation spectra.

These results suggest that invertible architectures can unify spectrum prediction and uncertainty-aware candidate generation within one end-to-end model.