From Natural Language to Executable Optimization Workflows: A Knowledge-Graph-Driven Framework for Multidisciplinary Optimization

Abstract

Early-stage engineering design requires balancing competing performance, cost, and feasibility criteria across interacting disciplinary models. However, deploying multidisciplinary optimization (MDO) frameworks remains difficult because translating high-level requirements into executable optimization workflows often requires substantial manual effort in workflow configuration, tool integration, and dependency management. This paper presents a knowledge-graph-grounded framework that automatically translates free-form natural-language requirements into validated and executable MDO workflows. A locally hosted large language model extracts structured optimization intents, while a domain knowledge graph maps these intents to canonical variables, retrieves metadata, and generates dependency-consistent execution plans. Candidate designs are evaluated through orchestrated Python-based model execution with workbook reuse, and a population-based genetic optimizer searches the bounded decision-variable space while enforcing feasibility through penalty-based ranking. Transparency is supported through parsed intents, confidence scoring, and deterministic post-validation. The framework is demonstrated through a vehicle concept optimization case study for early-stage design exploration.