Transformer Encoder vs. Mamba SSM: Lightweight Architectures for Machining Stability-Induced Surface-Quality Categorization

Recent advances in large language models (LLMs) have revolutionized many domains; however, their adoption in manufacturing remains limited. This article explores the potential of state-of-the-art AI methodologies for chatter-induced surface-quality categorization in machining. While LLMs typically demand extremely high computational resources (FLOPs), AI systems for chatter recognition must instead be designed for efficient classification or regression on structured sensor signals—requiring compact architectures capable of approximately—0.001 s inference. To address this, the present study applies and compares two representative architectures: the transformer encoder, an attention-based model, and the Mamba SSM (State-Space Model), a non-attention-based state-space alternative. Two case studies are conducted to evaluate their performance: one using machining-dynamics-based simulation data, and the other employing actual experimental measurements. Finally, the advantages and limitations of these cutting-edge AI approaches are critically discussed, emphasizing their suitability and practical challenges for real-world manufacturing deployment.