Doc 2 Model

NASA-MSTAR(Doc2Model)

Teaching AI to Read Like an Engineer: FSU and JPL Team Up on Doc2Model

NASA wants its engineering to run on models, but most projects still run on documents. Teams supporting the Artemis campaign, and the suppliers behind them, still hand-write concepts of operations, requirements specifications, interface control documents and verification plans.

Model-Based Systems Engineering (MBSE) promises automated consistency checks, traceability and analysis. But building rigorous models up front is expensive, and MBSE experts are scarce. So many projects stay put.

This project funded at the Intelligent Systems Lab of Fayetteville State University (FSU) by NASA’s MUREP Space Technology Artemis Research (M-STAR) program, in partnership with NASA's Jet Propulsion Laboratory (JPL), asks a simple question: what if AI could build the model from the documents a team already has?

What Doc2Model does

The project is called Doc2Model. It uses generative AI to reverse-engineer existing engineering documents into formal, methodology-conformant models, and then regenerates the documents from those models. The pipeline runs in five stages:

1. Discover the vocabulary. A language model reads a project's documents and finds the implicit terms engineers use: entity types, properties, relations and constraints, each linked back to the source text.
2. Curate and map it. Engineers review that vocabulary and map it to a target methodology written in the Ontological Modeling Language (OML), part of JPL's openCAESAR framework.
3. Extract the knowledge. The AI turns document content into OML models, but every change goes through validated tools that enforce the methodology's rules.
4. Check the model. A linter checks syntax, a validator checks methodology rules, and a reasoner checks logical consistency.
5. Regenerate the documents. The validated model produces the documents again, and the team measures how faithfully they match the originals.

That last step is the clever part. Teams keep working in documents they trust while quietly gaining a model underneath, an on-ramp to MBSE rather than a cliff.

Why "method compliance" matters

Large language models are fluent, but fluency is not correctness. In the team's earlier MDE Intelligence 2026 paper, an unassisted frontier model wrote syntactically perfect OML yet achieved zero method compliance on complex tasks. The authors call this "hallucination-by-commission": output that looks right and breaks the rules.

Their fix is architectural. The LLM interprets intent and orchestrates, but it never edits the model directly. Changes are routed through Model Context Protocol (MCP) tools that call the same validated APIs as the OML IDE, with methodology rules encoded as SHACL shapes and enforced during authoring. The team will also compare low-cost, locally hosted open-source models against leading proprietary ones, and release the benchmark, prompts and code publicly.

The people behind it

Dr. Sambit Bhattacharya, Principal Investigator. Professor of Computer Science at FSU, and founding Director of its Intelligent Systems Laboratory. 

Dr. Maged Elaasar, JPL Technical Collaborator. A Senior Technology Researcher at JPL and the architect of openCAESAR, the ontological MBSE framework used on JPL-led NASA missions for over a decade. He has built a GenAI modeling assistant for OML, and will guide methodology, co-design Artemis validation cases and mentor students.

Dr. Steven Jenkins, Consultant. Retired in 2021 as a Principal Systems Engineer at JPL, where he was Chief Engineer of the Integrated Model-Centric Engineering initiative. He introduced semantic technologies to JPL systems engineering, laying groundwork for CAESAR, and prototyped a model-based generator that produced hundred-page documents in seconds. He will review vocabularies, validation rules and document generation.

Ms. Sokhna Amar, graduate student and first author of the MDE Intelligence 2026 paper leads the student team. 

Dr. Adrienne A. Smith and Dr. Rebecca Zulli Lowe, Independent Evaluators. From Cynosure Consulting, they will independently assess progress, including re-running the published scoring to confirm reported results.

A full-time postdoctoral researcher, still to be recruited, will lead day-to-day technical work and coordinate student teams.

Building the next digital-engineering workforce

Doc2Model is as much about people as software. Each year it funds graduate and undergraduate researchers, and places students in summer internships at NASA JPL or another NASA Center. Students will build and test the pipeline on real aerospace documents, run experiments on the lab's GPU clusters, and present results to JPL practitioners. FSU will also add a two-module "AI-Assisted Model-Based Systems Engineering" sequence to its AI and capstone courses, reaching at least 20 students a year and released openly for other universities.

If it succeeds, document-heavy projects and suppliers would get a low-cost way to start using MBSE without abandoning how they work today. And a new cohort of students would be ready to carry that work forward.