Dual-graph computable knowledge engineering for CAD-to-CNC macro synthesis

Abstract

Large language models (LLMs) can synthesize G-code, but autonomous CAD-to-CNC macro-generation must preserve geometry, process constraints, variable roles, equations, loop bounds, coordinate mappings, and controller syntax. This paper presents a dual-graph computable knowledge engineering method in which typed knowledge objects are selected by a semantic relation graph and converted by a computational dependency graph into acyclic variable-expression-loop dependencies. A seven-layer Macro-Structured Parametric CNC Code (Macro-SPCC) schema and Macro-Program Intermediate Representation (MPIR) organize the evidence, while deterministic modules perform evaluation, geometric checking, and FANUC Macro B compilation. Across 12 tire-mold lettering-insert tasks from six DXF cases and two toolpath configurations, the method achieved 91.7% accuracy, 50.0 percentage points above Text-RAG, and a 100.0% compilation pass rate. VERICUT confirmed continuity, endpoint reachability, and safe retraction, demonstrating a traceable, computationally closed, and executable synthesis process.