美光科技不只是在谈论人工智能(AI). 在自己的制造过程中使用数据分析和人工智能, 该公司确实把钱花在了说到做到的事情上, demonstrating the value to businesses of the technologies Micron enables with its next-generation memory and storage solutions. The benefits are many, including higher yields, a safer working environment, 提高效率和可持续发展的业务.

The enterprise’s factories produce memory technologies on silicon wafers through a highly complex and precise process. 出错和浪费的可能性很高. 但数据和人工智能正在帮助降低这种可能性. 当依靠人类的警惕性来发现和跟踪缺陷时, 机械问题和其他问题领域, the organization lost time and money — losses that can be avoided with today’s sophisticated technologies.

Complex manufacturing process

Silicon manufacturing is an extremely complex process, taking months and involving some 1,500 steps. 美光在整个过程中使用了复杂的人工智能来提高准确性和覆盖范围.

“我们在这里打造了一些完全与众不同的东西,” explains Koen de Backer, 美光智能制造和人工智能副总裁. “我们看到的准确率要高得多. 我们现在可以以两倍的速度推出沙巴体育结算平台，同时将生产率提高10%. It’s truly been transformative. You could say it’s a killer app.”

Wafer creation

The process starts with silicon. Wafers, 用作计算机芯片的基础, are made from silica, a type of sand, 哪些必须过滤提炼成99.999% purity. 这种电子级硅被熔化并压缩成锭, 它们被切成极薄的0.67mm thick — wafers.

对晶圆片进行抛光以去除切割留下的痕迹, 涂上一层薄薄的抗光材料, and etched with the design of the circuitry they will be supporting using a process similar to photography. 电路越复杂，晶圆上印上的图像就越多. 这个过程一层一层地进行, 每层都分开处理，或者用电离等离子体轰击, a process known as “doping,” or bathed in metals. The finished wafer is then coated with a thin protective film before being tested to ensure that it works as intended.

The manufacturing process takes place in sterile fabrication rooms (called cleanrooms) designed to prevent even the tiniest speck of dust from falling on the pristine wafers. But damage does occur. 易碎的晶圆片可能被划伤, nicked or punctured, 或者在保护膜下形成气泡.

通常，这些缺陷是微小的——肉眼完全看不见. Even when they are visible, people scanning the 30 to 40 photos captured of each wafer during the photographic imaging process can overlook defects due to eye fatigue or momentary inattention. 一眨眼，他们就错过了瑕疵.

When problems aren’t caught until the test phase, a lot of time and money has already been wasted. 造成缺陷的问题很可能影响不止一块晶圆片，甚至可能是数千块.

在生产环境中也可能出现其他问题. Parts wear out; pipes leak or drip hazardous chemicals onto products or people. 及早发现并纠正这些问题是必要的. 停工成本非常高，会导致收入损失和生产力损失. 考虑到半导体制造的复杂性, 在恢复过程中花费的大量时间可能会使实际成本达到数百万美元. 更重要的是，与工人受伤相关的风险是无数的. Finally, Micron’s commitment to sustainability demands that the process become as energy efficient as possible.

检测沙巴体育结算平台和机械中的问题对于提高生产效率至关重要, effectiveness and safety. Unfortunately, to err is human, 即使是最训练有素的人也不一定能看到, 听到或感觉到非常微小和微妙的迹象，表明有些事情是错误的.

人工智能技术, however, 能在很短的时间内以极高的精度完成这些任务吗. 美光从570多家公司收集了pb级的制造数据,这些资源被添加到美光的云数据库环境中.

Image analytics

美光人工智能制造的基础是图像分析. “图像在半导体制造过程中非常强大,” explains Koen, “你可以分析过程中每一步的详细图像.”

“By analyzing every stage,” he continues, “我们可以快速识别发生的任何偏差——所有这些都是完全自动化的. 这种分析涵盖了所有方面——前端、组装和测试.”

In addition to images, 美光同样采用视频分析来消除组装和测试中的质量问题. 你可能会认为视频的数据量太大，不实用. 然而，美光再次使用人工智能来识别需要分析的关键位置. 人工智能启动和停止视频流，只捕捉关键的过程, keeping data size in check.

成像和视频尤其有效，因为晶圆缺陷有多种形式. For the most part, they fall into one of a few common categories: tiny holes near the wafer’s edge or scratches and bubbles in the outer film. Micron’s AI systems use “computer vision” technology to spot these defects on the images the photolithographic cameras capture as they etch circuitry onto the wafers during manufacturing.

工程师可能会指示系统扫描晶圆片边缘的小点(孔), for instance, 或者对于连续的或轻微的断线(划痕), 或者系统可能会寻找颜色变化，导致深色或浅色斑点或图案. 其中一些缺陷几乎可以实时发现, 该系统在拍摄图像后几秒钟内发出警报. Other defects might be discovered during secondary scans minutes after the photographs are stored. All these processes rely on the AI system’s use of two million images stored in the database environment for comparison and contrast.

事实证明，其结果比工程师的评估准确得多, 由于人工智能计算机视觉具有更高的精度和高效率. 最重要的是，工程师现在可以专注于问题和数据收集.

并采用美光的AI自动缺陷分类(ADC)系统, 技术人员和工程师不再需要手动对晶圆缺陷进行分类. 相反，AI-ADC每年使用深度学习对数百万个缺陷进行排序和分类. 美光利用当今最新的成像技术创造了这个系统, including neural networks, described as a biologically inspired programming paradigm that enables a computer to learn from observational data.

这种形式的机器学习根据缺陷对图像进行分类, placing them in discrete clusters. Not only does this process help engineers to discover what went wrong during manufacturing for an early fix that avoids more defects, 但它也使人工智能系统能够自己发现缺陷，并在每次迭代中改进结果.

Acoustic listening

而人工智能成像是制造过程的核心, 美光还采用声学聆听来先发制人. 通常，“不正常”的声音表示一个磨损的部件或即将崩溃.

Micron’s AI systems are listening for anomalies in our factory machinery via audial sensors installed strategically near robotic actuators or close to pumps. 这些麦克风记录了几个星期的正常活动, and software converts the detected frequencies into graphs or charts depicting the sounds as visual data. 当出现新的音高或频率时，系统会发出警报. 通常，它甚至可以辨别出异常的原因.

搜索这些庞大的数据库非常耗时. 然而，当一台机器有故障的危险时，工厂经理需要立即知道. 将数据发送给充满GPU的人工智能系统, accelerators, 更重要的是，提供快速的内存和存储, 智能的结果比基于cpu的系统更快. All these AI systems with hundreds of thousands of GPU cores and memory working simultaneously and synergistically can refine their results in the blink of an eye with little or no human intervention. Plus, they can improve their diagnostics with each iteration, similar to how the human brain works.

Thermal imaging

并不是所有的故障都会产生噪音——在制造环境中，沉默可能是致命的. 在许多情况下，温度反而发生了变化. Until recently, the only way to detect a surge in temperature was to see a red glow, sparks or smoke. By the time these appeared, 问题已经进入了危险区, 核电站需要疏散工人.

So, 除了图像分析和听觉, Micron also uses thermal imaging, 哪个测量关键部件的温度.

“对变压器进行温度测量是防止过热的关键,” Koen explains, “早期检测可以决定是进行简单的维修还是更换整个设备, expensive piece of equipment.”

Finally, these AI sensors for imagery, 声音和温度实现了美光对可持续发展的承诺. “这些传感器在提高质量和效率方面表现出色, but also for sustainability,” Koen adds. “它们提供了精细的能源计量，从而实现了显著的使用和节能.”

The numbers

在美光，57万个传感器产生2.2.29亿个控制点的300万张晶圆图像. 所有这些都是每周通过人工智能模型运行的. 此外，还存储了34pb的数据，每天捕获30tb的新数据.

这种大规模的人工智能应用分析了产量分析中的创新数据科学应用, digital-twin planning, IoT and image analytics, 优化和高级算法, 过程自动化和移动应用程序.

• 制造工具可用性提高4%
• 劳动生产率提高18%
• 新沙巴体育结算平台上市时间缩短50%
• 22% reduction in product scrap
• 解决质量问题的时间缩短50%

^* 这些改进是基于2016年至2020年收集的美光内部数据和分析

And the benefits of data analytics and AI extend beyond the fab to every aspect of Micron’s operations: sales and marketing, human resources, business operations, research and development, and more.

科恩表示:“这关乎企业的转型，而不仅仅是车间. “我们可以将这些技术和方法应用到公司的所有业务流程中.”

Ecosystem partnerships

除了优化内部制造流程, 美光也直接与供应商合作, 向他们提供沙巴体育结算平台的详细反馈，以确保最佳的能源效率. With these suppliers, Micron coordinates DIMS (data ingestion into Micron systems) so that ingestion happens at the highest frequencies. 美光工程师实时监控这种吸收. 同样地，修正和优化也不断地发生在极细粒度的级别上.

In addition, working with our suppliers, 美光使用遥测数据来测量我们的沙巴体育结算平台在他们的数据中心的效果. This data, combined with internal data, 支持实时协作，以针对特定工作负载改进沙巴体育结算平台.

我们还密切监视模型的性能. Using AI on incoming data and retraining allows engineers to focus at a higher level on automating the machine learning flow. (Otherwise, 永远都不会有足够的数据科学家来跟上, 他们会追踪已经发生的事情.)

These initiatives are backed by an internal data science academy and continued investment in in-house data scientists, engineers and solution architects. These resources, 还有我们的公民数据科学模型, 使职能专家能够使用人工智能驱动的工具和见解.

Industry leadership

Today, Micron combines a rich heritage of core process knowledge with the unparalleled efficiencies of AI. Data experts have created large yield management platforms used by 6,000 people across the company. At the same time, dedicated teams focused on day-to-day yield optimization are building new prototypes in fast integration cycles. 这些原型经常用于优化主要平台.

与此同时，结果不言自明. The dedication of Micron team members and AI-powered manufacturing processes enabled our 1α (1-alpha) node DRAM and 176-layer NAND to hit the highest yields in Micron’s history. And the industry-leading 1β (1-beta) DRAM and 232-layer NAND achieved mature yields faster than any other Micron technology.

美光正在优化人工智能如何改变制造业. Far from taking everyone’s jobs, this new technology is augmenting and empowering teams who no longer focus on acquiring data and doing multiple base analyses. 现在他们可以专注于自己擅长的领域——创新，开发行业领先的沙巴体育结算平台.