Smart data adapter: A hybrid pattern-LLM approach

Adopting digital twin technologies in small andmedium-sized enterprises (SMEs) is often hinderedby heterogeneous, poorly documented productiondata. Current semantic type detection approachesrequire massive labeled datasets making themimpractical for resource-constrained SMEs. Wepropose a zero-shot hybrid framework combiningpattern-based classification with selective largelanguage model (LLM) reasoning formanufacturing-specific data types. The two-stagearchitecture uses rule-based patterns for high-confidence cases, forwarding ambiguous columnsto a multi-step LLM reasoner. Evaluation on fourmanufacturing datasets shows the hybrid approachachieves weighted F1 within 7 24 percentagepoints of pure LLM classification performancewhile reducing LLM invocations by 39% onaverage. Processing time decreased by up to 2.8×.Our framework addresses a critical gap: automated,computationally efficient data type recognition formanufacturing SMEs without requiring trainingdata, contributing to automated simulation anddigital twin construction.