🤖 An LLM-Agentic Workflow for Data-Driven Modeling

From Toy Image Reconstruction to Cu–Mg CALPHAD Assessment

---

🌐 Overview

This repository provides the code implementation for our paper:

“An LLM-Agentic Workflow for Data-Driven Modeling: From Toy Image Reconstruction to Thermodynamic Modeling” preprint available at: https://www.researchsquare.com/article/rs-8574739/v2

We introduce Auto-DDM (data driven modeling), a novel hybrid optimization system that integrates large language model (LLM) agents into a multi-objective evolutionary algorithm. The method significantly improves the quality and efficiency of continuous optimization, particularly in data-scarce, high-dimensional domains.

---

🚀 Key Features

• LLM-Embedded Optimization
  GPT-driven agents participate directly in the optimization loop to intelligently modify parameters and control diversity.

• Outperforms Human Experts
  Achieves superior results compared to expert-designed dataset weighting in a 22-objective CALPHAD task.

• Two Intelligent Agents

  • Repair Agent: Proposes edits or ε-threshold changes based on population statistics.
  • Diversity Agent: Actively prevents early collapse by perturbing over-converged parameters.

---