脈衝雷射加工表面形貌分類系統建構與分析

Abstract

雷射加工技術因具備非接觸式、高加工精度、可改變材料表面特性及環保優勢，近年來在製造領域中扮演著日益重要的角色。隨著對加工後表面品質要求的提升，現有雷射加工技術在加工參數與表面形貌對應關係之資料庫建置方面，仍存在明顯不足。目前市售雷射系統多僅提供表面加工範例，卻未公開完整加工參數資訊，導致使用者需仰賴大量試誤調整，耗時且效率低落。此外，現有研究多集中於特定表面現象，如超疏水性或摩擦性，缺乏針對不同加工條件下全面性表面形貌數據之系統性整合。當前雷射加工表面形貌資料具有以下挑戰，包含形貌變異性高、多重加工參數交互影響，缺乏統一形貌標籤等。 本研究涉及奈秒、皮秒以及飛秒，涵蓋廣泛之脈衝寬度以探討多樣的加工表面形貌，並導入無監督學習進行分類，與常見之先定義標籤後探索表面參數之研究有所不同。提出以掃描間距為條件的兩階段分群策略，歸納出六種語意明確的表面形貌類型，並建構基於自編碼器特徵之無監督式分類模型，有效減少主觀判斷誤差。透過多模態視覺化分析，發現各類型與雷射參數間具高度對應性，並揭示加工參數間之綜效關係與操作窗口，最終建立具語意解釋性與回溯能力之分類系統，可應用於加工優化與品質預測。

Laser processing technology has become increasingly vital in modern manufacturing due to its non-contact nature, high machining precision, capability to modify surface properties, and environmental advantages. With rising demands for post-processing surface quality, current laser processing technologies still face significant limitations in establishing comprehensive databases that correlate processing parameters with surface morphologies. Most commercial laser systems only provide example cases of surface processing without disclosing complete parameter information, forcing users to rely heavily on time-consuming trial-and-error adjustments. Furthermore, existing studies predominantly focus on specific surface phenomena, such as superhydrophobicity or frictional properties, and lack systematic integration of surface morphology data across diverse processing conditions. Presently, surface morphology data in laser processing presents several challenges, including high morphological variability, complex interactions among multiple processing parameters, and the absence of standardized surface morphology labels. This study investigates nanosecond, picosecond, and femtosecond regimes, covering a broad range of pulse widths to explore diverse laser-induced surface morphologies. Unlike conventional approaches that predefine surface labels before exploring processing parameters, this work adopts an unsupervised learning framework for classification. A two-stage clustering strategy, conditioned on hatch distance, is proposed to derive six semantically distinct surface morphology types. An unsupervised classification model based on autoencoder-extracted features is developed, effectively reducing subjective labeling bias. Through multimodal visual analyses, strong correlations between morphology types and laser parameters are revealed, along with synergistic interactions among processing parameters and their operational windows. Ultimately, this study establishes a semantically interpretable and traceable classification system, which can be utilized for process optimization and quality prediction.