Title: DiffUNet^2: Bidirectional Prediction, Probabilistic Generation and Collaborative Visual Discovery for Scientific Data

Abstract

Analyzing and reasoning about scientific phenomena relies heavily on modeling temporal evolution. However, conventional machine learning approaches typically offer deterministic forward predictions, which fail to account for multiple plausible outcomes and seldom facilitate backward reasoning. This limitation restricts their utility within practical scientific workflows. To address these challenges, we introduce a framework that combines interactive visual analytics with diffusion-based generative modeling to enhance scientific exploration.

Central to this framework is DiffUNet^2, a conditional diffusion model designed to perform bidirectional, any-to-any generation across time. This model effectively captures the distributions of plausible system evolutions. Leveraging this architecture, we developed an interactive system that facilitates branching timeline exploration, user-guided state editing, and navigation through probability space. These features empower scientists to actively investigate alternative hypotheses rather than merely observing static predictions.

We assessed the model’s performance using five datasets spanning various scientific fields, validating both its predictive accuracy and the quality of its probability-space ensembles. Furthermore, through collaboration with domain experts, we demonstrated the approach’s efficacy in supporting real-world scientific temporal data analysis workflows. By merging generative modeling with visual interaction, our method transforms generative models into powerful tools for hypothesis-driven analysis, allowing scientists to interactively explore system dynamics.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC