FMSG: Cyber: Cybermanufacturing of Wide-Bandgap Semiconductor Devices Enabled by Simulation Augmented Machine Learning

半导体行业是最大的制造业之一，年收入接近5000亿美元。通过多个工艺步骤在大直径晶片上制造半导体器件。成品率是决定半导体制造成功的关键指标。目前的良率管理实践依赖于最小化晶片材料的不均匀性，最大化每个步骤中的工艺控制，以及基于领域专业知识对整个晶片应用必要的工艺适应。然而，新兴半导体器件的制造良率，由于不太成熟的材料和工艺，宽带隙（WBG）器件在代工厂中仅占50-80%。虽然WBG设备在电动汽车、数据中心、5G通信和电网等应用中得到了快速采用，但其制造的产量有限已成为一个日益严重的问题。这个未来制造种子基金（FMSG）网络制造项目建议通过在物理仿真增强机器学习（ML）框架指导下的每个工艺步骤中基于芯片或器件（而不是基于晶圆）的适应来实现半导体器件的自预测和自适应网络制造。在这种目前尚不存在的半导体网络制造中，在每个工艺步骤中应用几何和设计中的器件到器件的适配，以智能地补偿固有材料特性和历史工艺步骤的可变性。该种子基金将使用WBG功率二极管的小规模制造作为示范工具，以建立与自适应半导体制造中ML集成相关的知识库。新的制造模式可能会在ML和半导体的交叉点形成新的产业。该项目还提供了一个独特的场所，以培养未来的技术人员，解决ML引导的半导体制造中的跨学科问题的能力。这个跨学科项目将用于支持本科生的研究活动和K-12学生的推广活动。 该种子基金的目的是确定和解决与半导体网络制造相关的基本知识差距，使用垂直氮化镓功率二极管的小规模制造作为示范车辆，这是一种新兴的WBG器件，用于电动汽车和电网的功率应用。该项目的智力价值植根于半导体器件制造的全新理念，即，由分析和预测ML模型产生的管芯到管芯、器件到器件的适配。为了实现这种新的制造模式，该项目将重点解决以下问题：（a）将探索新的数据框架，以开发适用于物理电子设备的ML模型。在成本和时间方面昂贵的实验设备数据将通过使用技术计算机辅助设计模拟的物理模拟数据增加1000 - 10000倍。(b)将针对正向过程（根据给定的材料/器械参数集预测器械性能指标）和反向过程（针对给定器械特性、测量的历史过程步骤参数和设计目标推导未来过程参数）探索创新ML模型。(c)所提出的框架将通过在测试车辆上的试验性制造进行实验性证明，该未来制造项目由工程局的ECCS和CMMI部门以及数学和物理科学局的CHE部门共同资助。该奖项反映了NSF的法定使命，并被认为值得支持通过使用基金会的知识价值和更广泛的影响审查标准进行评估。

Semiconductor industry is one of the largest manufacturing industries with annual revenue approaching $500 billion. Semiconductor devices are manufactured on large-diameter wafers through multiple process steps. Yield is a key metric determining the success in semiconductor manufacturing. The current practice of yield management relies on minimizing the wafer material non-uniformity, maximizing the process control in every step, and applying necessary process adaptions to the entire wafer based on domain expertise. However, the manufacturing yield of emerging semiconductor devices, e.g., wide-bandgap (WBG) devices, is merely 50-80% in the foundry, due to less mature materials and processes. While WBG devices are gaining quick adoption in applications like electric vehicles, data centers, 5G communications, and power grids, the limited yield of their manufacturing has become an increasingly serious concern. This Future Manufacturing Seed Grant (FMSG) CyberManufacturing project suggests the self-predictive and self-adaptive cybermanufacturing of semiconductor devices implemented through die- or device-based (instead of wafer-based) adaptions in each process step guided by a physical simulation augmented machine learning (ML) framework. In this semiconductor cybermanufacturing, which does not exist today, device-to-device adaptions in geometrics and designs are applied in each process step to intelligently compensate for the variability in inherent material properties and historical process steps. This seed grant will use the small-scale fabrication of WBG power diodes as a demonstration vehicle to establish the knowledge base related to the integration of ML in adaptive semiconductor manufacturing. The new manufacturing paradigm can potentially lead to the formation of new industries at the intersection of ML and semiconductors. This project also presents a unique venue to train future technicians with the capabilities of tackling interdisciplinary problems in ML-guided semiconductor manufacturing. This interdisciplinary project will be utilized to support undergraduate research activities and outreach activities for K-12 students. The objective of this seed grant is to identify and address the fundamental knowledge gaps related to the semiconductor cybermanufacturing, using the small-scale fabrication of vertical gallium nitride power diodes as a demonstration vehicle, which is an emerging WBG device for power applications in electric vehicles and power grids. The intellectual merits of this project are rooted in the fundamentally new philosophy for semiconductor device manufacturing, i.e., the die-to-die, device-to-device adaptions produced by analytic and predictive ML models. To realize this new manufacturing paradigm, this project will focus on tacking the following problems: (a) New data frameworks will be explored for the development of ML models applicable to physical electronic devices. Experimental device data, which are expensive in terms of cost and time, will be augmented by physical simulation data by 1,000-10,000 times using the Technology Computer-Aided Design simulations. (b) Innovative ML models will be explored for the forward process (predict device performance metrics from a given set of material/device parameters) and inverse process (deduce future process parameters for the given device characteristics, the measured historical process step parameters, and the design objectives). (c) The proposed framework will be experimentally demonstrated through pilot manufacturing on the test vehicle, and the final yield enhancement will be characterized and evaluated.This Future Manufacturing project is jointly funded by the Divisions of ECCS and CMMI in the Directorate of Engineering and the Division of CHE in the Directorate for Mathematical and Physical Sciences.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.