Munich-based Infineon Technologies AG, manufacturer of semiconductor solutions, microcontrollers, sensors and other chips, deals with the smallest components. This makes it difficult for quality management to develop an efficient error analysis that detects abnormalities even in microscopic conductor tracks. Even a small grain of dust can make a chip unusable. In the iFAME project (“Infineon Failure Analysis Image Engine”), the group has now implemented the digitization of error analysis.
Finding defects in semiconductor production can take two months and more using the traditional approach. In order to digitize the process, the project team first collected a large number of analysis reports from the areas of development, production and customer applications with information on the fault characteristics, problem description and cause of the fault. Such data has been in the group for 15 years, but a viable solution to consolidate and analyze this unstructured data was lacking.
That’s where the iFAME application comes in, a solution for combined image and text searches in a global database. It enables the company to save up to 80 percent in error analysis and has a hit accuracy of over 90 percent. iFAME laid the foundation for comprehensive digital quality management at Infineon worldwide – reason enough for the team to apply for the “Digital Leader Award” (DLA), which the IDG publishing house awards together with the NTT Group.
In order to steer the project in the right direction, a different perspective was a prerequisite at Infineon: Quality is not just an issue of the quality department, it concerns everyone in the group, including the IT organization – and continuously. In general, IT is viewed differently in the Group today than it used to be. Your most important task now is to convert digital innovations into concrete benefits. The cross-divisional project iFAME serves as a role model because global quality management works in an exemplary manner with IT.
Today, test engineers can use the “Search Engine” iFAME to enter search queries for error identification using images and text. A multi-layered neural network is used for image search, which extracts the abstract features of error analysis images and uses algorithms to examine them for similarities or deviations. In addition, there is a semantic search based on deep learning that interprets the meaning of words (context and syntax). This means that meaningful search results can be achieved without the terminology used in the database being exactly known.
A third element of iFAME is database indexing: In order to quickly obtain relevant and accurate search results, Infineon uses a modern indexing algorithm that has been tailored to very large amounts of data. The data is sorted and sorted in advance so that pre-structured information is available at the start of the search.
The global implementation of iFAME was done in an agile way. A minimum viable product (MVP) was created in short iteration cycles, which was further developed and optimized in sprints. Methods such as Proof of Concept and Proof of Value ensured that the requirements in the areas of scalability, security and user experience were constantly compared. With well-planned change management measures, the target groups were prepared for the upcoming changes by iFAME.
In order to master the challenges of image and text search at iFAME, image and text data were converted into a vector space representation and combined with classification and clustering algorithms, linear transformations and machine learning-based ranking models. IT ensured the required computing capacity of the deep learning models used through the GPU-supported computing power of its own Infineon Compute Farm. The project team describes the provision of online models with high parallelism and low latency as a “complex requirement for a multi-deep learning model backend”. It requires several levels of profile creation and tool testing.
Thanks to modern technologies such as Elastic Search and CEPH (as a distributed object-based storage solution), search results can now be displayed in the database with more than 1.7 million images and around 78,000 reports in less than five seconds on average. An important basis for this is the “Small World Graph” indexing algorithm, developed from scientific approaches, with low latency and fine-tuned hyperparameters.
The scalability of the search was ensured thanks to the expert knowledge of the participants as well as the experience with DevOps approaches and the Openshift ContainerTechnology Stack from IBM / RedHat, although not only the number of error analysis reports is constantly growing, but also that of the users. In addition, the Munich-based company successfully expanded its internal AI skills.
Today, iFAME at Infineon is a successful, globally used solution that helps the more than 700 test engineers to carry out complex error analyzes faster and to identify possible solutions. This enables Bayern to answer customer queries much faster. They rely on a database with more than 1.7 million images and around 78,000 reports. Furthermore, iFAME leads to a worldwide network of previously local laboratories for error analysis and creates a worldwide virtual network of experts.