As the semiconductor packaging technology evolves in the “More than Moore” era, many advanced processes have presented challenges in manufacturing. Often, challenges in precision die bonding, for example, have led to trade-offs in yield, throughput, and ultimately manufacturing cost. Precision alignment has become one of the key capabilities for several advanced technologies for heterogeneous integration, such as 2.5D/3D, optical assembly, etc.
For advanced packaging, over the years, as the package pitches continuously decrease for high-density designs, die bond alignment at finer pitches, such as 10µm or below, has become a key challenge area. For flip-chip die, for example, there is a cost tradeoff with high accuracy on most bonding equipment —- the tighter the pitch, the slower the die bond alignment process.
Limitations on the die alignment precision often can impose limits on the device integration density that can be achieved using 3D stacking.High-density fan-outis another process that requires high-precision die bonding. Once the requirement for die alignment precision goes below 1μm, challenges in yield and throughput become even more severe. For hybrid bonding, for example, the alignment precision of the equipment can be a limiting factor for the interconnect pitch, thereby limiting the realization of the design requirements in certain situations. Theemergence of chipletspresents even greater challenges for volume manufacturing.
High precision alignment is also a critical capability for transceivers, fiber optics, MEMS, LiDAR, LED, optical imaging, and others. These important applications demand high precision alignment and high throughputsimultaneously, further pushing the capacity limit of the equipment. For optical assembly, for example, lens attach, transmitter placement, receiver position, bond line thickness (BLT), etc. are all critical parameters for process yield and product performance.
机器学习和人工智能（ML/AI）可以成为帮助应对这些挑战的非常有效的工具。
Machine Learning is focused on finding patterns in data and using those patterns to make predictions. ML/AI can have many important applications in semiconductor packaging and electronics manufacturing. These range from supply chain management, new product introduction (NPI), production planning, traceability, to predictive maintenance, productivity improvement, and process improvement, yield improvement, among others.
One example is an ML/AI platform that utilizes real-time machine data and operational data (such as actual placement positions, cycle time, etc.) to develop ML models right there at the machine (i.e. at the “edge”). Easy interfacing with the machine data sources (such as existing sensors, actuators, PLC, log data, etc.) via appropriate communication protocols, make available a large body of real-time data (such as acceleration, rotation, conveyor speed, gripper position, bonding head motion, epoxy dispense pressure, and other real-time parameters of various machine parts),. which are important for machine and process performance. Real-time data ingestion and multi-variable data processing, in combination with data transformation, stitching, and filtration, provides a robust foundation for mass data virtualization for building ML models (Figure 1).
The platform collects data in milliseconds and builds ML models in days and weeks, in an actual manufacturing environment without interrupting production. The ML models, through continuous refinement in real-time, are used for predictive maintenance (by anomaly detection) (Figure 2) and more importantly, for adaptive control of the machine itself (Figure 3). These capabilities have been leveraged by several large global enterprises for semiconductor fabs, advanced packaging, fiber optics for transceivers, precision placement for SMT, etc. Significant improvements have been achieved to improve the alignment accuracy and process yield while reducing cycle time and machine downtime (Figure 4).
与此同时，可视化（图 5）和可追溯性（图 6）可以从现场/生产线级一直延伸到机器传感器级，其粒度是以前无法想象的，从而为优化和持续改进提供前所未有的洞察力和智能。这种解决方案（或根据需要的 "混合 "解决方案）还能确保数据安全和低延迟。由于该平台与机器无关，因此可以部署到工厂生产线上的各种机器上。由于该平台与机器无关，因此可以部署到工厂生产线上的各种机器上。这一点尤为强大，因为上下游机器现在可以 "协作 "进行整体同步优化。
通过各种使用案例，我们已经证明，在半导体制造、先进封装、光学装配等具有挑战性的工艺中，可以利用边缘 ML/AI 平台来显著提高机器性能以及产量和吞吐量，特别是在精密对准方面。当然，它还可用于诊断复杂的多变量工艺问题，以及通过利用整个生产线各个工位的数据来缩短测试时间。
