CVEvolve: Zero‑Code Autonomous Discovery of Scientific Image‑Processing Algorithms

Scientific research increasingly generates massive, unstructured image data, yet many domain scientists lack the computer‑vision, image‑processing, and software‑engineering expertise required to develop custom analysis pipelines.

To bridge this gap, the Argonne National Laboratory (ANL) team created CVEvolve , a zero‑code autonomous agent framework that can autonomously discover algorithms for scientific data processing. The system requires no predefined problem architecture or fixed workflow templates and integrates code generation, result verification, and strategy optimization in a single stack.

The framework is built on LangGraph and centers on a large‑language‑model (LLM) agent that drives a generate‑tune‑evolve loop. Supporting tools include a file‑system manager, environment manager, image viewer, search‑state recorder, and web‑search adapters for arXiv, Semantic Scholar, and Tavily. Multi‑modal image handling middleware injects rendered images back into the dialogue for subsequent reasoning.

Three custom datasets were constructed for evaluation:

XRF image registration : 809 paired images (10‑30 px) with simulated translation, Poisson noise, and blur; 10 % held out for testing.

Bragg‑peak detection : a single diffraction image split into development and holdout sets, annotated for peak locations.

High‑energy diffraction microscopy segmentation : five images with manual masks, plus two holdout samples.

XRF registration results showed baseline phase‑correlation error of 5.8 and brute‑force error of 1.25. After nine CVEvolve iterations the error fell to 0.43, and the best candidate achieved an error of 0.12—approximately eight times lower than the brute‑force baseline. A comparison with OpenEvolve after 500 iterations yielded a higher error of 0.23.

Bragg‑peak detection used F1, precision, and recall as metrics. Baseline scores were F1 0.298, precision 0.237, recall 0.400. The CVEvolve candidate selected at round 5 improved these to F1 0.788, precision 0.839, recall 0.743, reducing both false‑positive and false‑negative rates.

Diffraction‑image segmentation began with a baseline IoU of 0.37. After 16 search rounds CVEvolve produced a mask that closely matched the ground‑truth contours, successfully detecting most ring structures and Bragg peaks, with only minor missed regions in the outer area.

The study, titled CVEvolve: Autonomous Algorithm Discovery for Unstructured Scientific Data Processing and posted on arXiv, demonstrates that a no‑code, LLM‑driven approach can reliably generate high‑quality algorithms for diverse scientific imaging tasks, dramatically lowering the barrier for domain scientists. Future work aims to extend the framework to higher‑dimensional data and real‑time workflow optimization.