Scientific imaging often requires solving ill-posed inverse problems: recovering complete spatio-temporal fields from measurements that are indirect, noisy, and highly sparse. The lack of ground truth makes this especially challenging. A new arXiv preprint introduces NeuralDMD, a method aimed at exactly this setting.

The paper describes NeuralDMD as an interpretable neural representation of dynamics. While the abstract does not detail the architecture, the name suggests a combination of neural networks with dynamic mode decomposition, a common technique for extracting spatiotemporal patterns. The emphasis on interpretability is notable for applications where understanding the recovered dynamics matters.

Because this is a preprint, its claims have not yet been peer-reviewed. The abstract alone does not provide experimental results, so the reported performance remains to be evaluated.