LARA-HPC: Custom Agentic RAG for Thermodynamic Extractive Metallurgy

By Abdul Rahman · April 10, 2026 · Agentic AI

Building an intelligent Retrieval-Augmented Generation assistant to process and synthesize complex phase diagrams, CALPHAD datasets, and scientific literature.

The Data Bottleneck in Materials Engineering

When designing extractive metallurgy processes, engineers must cross-reference historical charts, thermodynamic values, and phase diagrams spread across centuries of literature. Finding specific binary or ternary phase boundaries, eutectic temperatures, or Gibbs free energy formulations can consume days of tedious document hunting.

To automate this workflow, I developed LARA-HPC, a custom Retrieval-Augmented Generation (RAG) assistant designed specifically for materials scientists.

Architectural Blueprint of LARA-HPC

LARA-HPC relies on a specialized agentic pipeline tailored for scientific text parsing:

1. Intelligent Chunking with Table/Equation Preservation: General purpose text-chunking breaks scientific tables or mathematical equations, making them unreadable. LARA-HPC utilizes markdown-focused parsing to keep thermodynamic tables intact.
2. Domain-Specific Embeddings: Standard embeddings struggle with words like martensite, solidus, or eutectoid. We augmented our vector search database with a custom metadata dictionary to prioritize metallurgy terminology.
3. Thermodynamic Verification Tools (Agent Tool-Use): To ensure accuracy, the AI LLM does not merely guess. It is equipped with custom Python computational tools (such as PyCalphad) to execute active equilibrium calculations on-the-fly when requested.

```python

conceptual RAG query enrichment

def search_thermo_db(query):

Searches specifically for phase boundary compositions

enriched_query = "phase boundary temperature liquidus solidus " + query return vector_store.similarity_search(enriched_query, k=3) ```

Enhancing Industry Workflows

For example, when optimizing the upgrading of low-grade iron ore at PCSIR (upgrading 36% Fe to 65% Fe), LARA-HPC synthesized 20 separate papers on indigenous mineralogy, delivering a structured breakdown of roasting temperatures and reduction parameters in under two minutes.

By shielding researchers from tedious searching and presenting synthesized, citation-backed physical data, LARA-HPC showcases the transformative power of Agentic AI inside the research laboratory.