| 摘要: | 隨著人工智能、網路科技的蓬勃發展以及5G射頻技術模組的需求下，硬體材料的LTCC低溫共燒陶瓷基板材料被推向歷史的高點。在小型化需 求下散熱及加工精度問題備受重視；陶瓷燒結後的硬度僅次於鑽石，所以後加工成本高昂，而燒結前的生胚則如粉筆、麵皮，所以陶瓷加工多採兩段式，生胚時做量大的粗加工，燒結後做精密加工；然而燒結後的收縮量造成極大的不確定性，致使燒結前加工精度在燒結後變得難以控制、再加上層間金屬電路的影響造成不均勻收縮，對尺寸精準度是一大考驗。
本論文提出在LTCC製程在生胚加工到燒結收縮後精度控制方法，整合自動光學量測、AI自動定位辨識，快速掌握QC結果，及時修正加工參數，透過深度學習原理，可以達成下列目標:
(1)可在同一批料即可快速比對燒結前、後尺寸收縮率誤差，及時修正生胚加工的參數設定，及時補償的概念，使最後目標尺寸中心點偏移縮小。
(2)從完工後汰除不良品的方式，進步到預測公差範圍，在超過公差前即提出暫停訊號，可使良率提升至4~5%以上，向100%目標逼近。
(3)相關連的任兩站，每一後工作站量測資訊立即回覆給前一站，透過深度學習，除即時優化精準度，更建立了完整資料庫，生產樣本越多可靠度越高，對於產品設計的優化、進料檢驗或製程改善都提供很好的工具。
;With the vigorous development of artificial intelligence, network technology and the demand for 5G RF technology modules, the LTCC low-temperature co-fired ceramic substrate material of hard materials has been pushed to a historical high. Under the requirement of miniaturization, heat dissipation and processing accuracy are highly valued; the hardness of ceramics after sintering is second only to diamonds, so the cost of post-processing is high, and the green embryos before sintering are like chalk and dough, so ceramic processing is usually two-stage. , Rough machining with a large amount of raw embryos and precision machining after sintering; however, the shrinkage after sintering causes great uncertainty, so that the processing accuracy before sintering becomes difficult to control after sintering, plus the interlayer metal circuit The effect causes uneven shrinkage, which is a test of dimensional accuracy.
This paper proposes a precision control method from green embryo processing to sinter shrinkage in the LTCC process, integrating automatic optical measurement and AI automatic positioning identification, quickly grasping QC results, and timely correcting processing parameters. Through deep learning principles, the following goals can be achieved:
(1)The size shrinkage error before and after sintering can be quickly compared in the same batch of materials, the parameter setting of green embryo processing can be corrected in time, and the concept of timely compensation can reduce the center point deviation of the final target size.
(2)From the method of eliminating defective products after completion, to the predicted tolerance range, the suspension signal is raised before the tolerance is exceeded, which can improve the yield to more than 4~5% and approach the 100% target.
(3)For any two related stations, the measurement information of each back station is immediately returned to the previous station, through the depth Learning, in addition to real-time optimization accuracy, a complete database has been established. The more production samples, the higher the reliability. It provides good tools for product design optimization, incoming inspection or process improvement. |
