Hidden in Plain Sight: Visual-to-Symbolic Analytical Solution Inference from Field Visualizations
Researchers' ViSA-R2 model turns 2D field visualizations into executable SymPy expressions with 8B parameters.
A research team led by Pengze Li has introduced ViSA-R2, a novel AI system designed to infer analytical solutions from scientific visualizations. The model tackles the "visual-to-symbolic analytical solution inference" (ViSA) problem: given a 2D plot of a physical field (like heat distribution or fluid flow) and minimal metadata, it must output a single, executable SymPy mathematical expression with all numeric constants correctly identified. This requires moving from pixels to precise symbolic mathematics, a fundamental but underexplored capability for AI in science.
The system is built on the open-weight Qwen3-VL 8B vision-language model backbone and employs a sophisticated, self-verifying chain-of-thought pipeline. It mimics a physicist's reasoning process: recognizing structural patterns in the visualization, hypothesizing a general solution family (ansatz), deriving specific parameters, and finally verifying the solution's consistency. To train and evaluate this approach, the team also released ViSA-Bench, a synthetic benchmark covering 30 linear steady-state physical scenarios with verifiable symbolic annotations.
In standardized evaluations, ViSA-R2 demonstrated superior performance, outperforming strong open-source baselines and the evaluated closed-source frontier vision models. The work represents a significant step beyond simple chart understanding, aiming to equip AI with the ability to recover the underlying governing equations—a core task in scientific reasoning and discovery.
- ViSA-R2 converts 2D field visualizations into executable SymPy expressions, recovering the underlying physics equations.
- Built on an 8B-parameter Qwen3-VL model, it uses a physicist-like reasoning pipeline for pattern recognition and verification.
- The team released ViSA-Bench, a new benchmark with 30 scenarios, where their model beat both open and closed-source VLMs.
Why It Matters
This enables AI to assist in scientific discovery by automatically deriving equations from data visualizations, accelerating research.