edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 26 SUMMARY Machine learning methods have already taken their first steps in FA applications. However, broad applications of AI in this domain require significant changes in the way semiconductor companies and tool vendors handle data in their software systems. Most of the data, at the moment, is stored unlabeled or with annotations in human-readable form only. For instance, in the latter case, FA engineers use some graphical editor to draw annotations on an image, whereas, in the former, the failure shown in an image is only described in text stored in a database. This data cannot be used by AI methods and significant investments are required to convert this data into machine-readable formats. Ideally, this digital transformation should be supported by the introduction of a general metadata storage and exchange format that can be used by internal FA systems and by the software of tool vendors. Easily accessible metadata can enable wide interoperability between tools and systems supporting daily FA operations. Thus, workflows become possible where annotations done in one program are easily imported and processed in another, like text or image editing software or AI applications. Interoperability on the FA operations level must also be raised to the organizational level, which might have a large impact since FA is connected to all other departments of the organization starting from development, quality management, manufacturing and up to marketing. The ontology-based approach discussed above is one of the possibilities, but others based on machine learning methods must also be investigated. A recent breakthrough in designing and training chatbots based on the generative pre-trained transformer (GPT) architecture[51] pushed the limits of information extraction, retrieval, and presentation to a new level. Open-source initiatives, like Open Assistant, (https://open-assistant.io) already aim for integration with third-party systems and dynamic information retrieval. All these advances take the next step toward general interoperability and information distribution among departments of a company, thus accelerating its overall learning ability development. Of course, training such large models requires huge amounts of data that, opposite to the training datasets of ChatGPT or Open Assistant, cannot be acquired from public sources, like the Web. For obvious reasons, organizations cannot share their images or texts online even in some anonymized form, like in the medical domain. Therefore, new approaches must be developed to guarantee the security and privacy of both the training process and resulting models. Modern methods such as differential privacy, secure multiparty computations, or homomorphic encryption[52] can be used to solve this problem from the technical perspective. However, companies should also invest in the organizational part by developing training infrastructures supporting all stages of machine learning operations (MLOps), like training, testing, deployment, and monitoring of models. CONCLUSIONS AI can significantly change the way FA labs perform daily operations as well as increase its impact on the global organizational level. Both symbolic and machine learning methods provide important contributions to the overall goal of FA digitalization and automation. However, the focus of existing applications mostly lies in solving only one problem at hand, like recognizing a failure on an image or performing an automatic scan test diagnosis. The development of global AI approaches, where all these components can cooperatively solve hard failure identification and localization problems, is still in its initial state. Therefore, AI-related initiatives, like the EDFAS roadmap, that coordinates multiple organizations in their activities, are of primary importance for the future of AI in FA. REFERENCES 1. M. Curtis: “Semiconductor Industry Leads in Artificial Intelligence Adoption,” Accenture Report, 2019. 2. S. Göke, K. Staight, and R. Vrijen: “Scaling AI in the Sector that Enables it: Lessons for Semiconductor-device Makers,” McKinsey Report, 2021. 3. A.M. Turing: “Computing Machinery and Intelligence,” Mind, LIX(236), 1950, p. 433–460. 4. P. McCorduck and C. Cfe: Machines Who Think: A Personal Inquiry into the History and Prospects of Artificial Intelligence, CRC Press, 2004. 5. S. Russel and P. Norvig: Artificial Intelligence: A Modern Approach, Pearson, 2020, ISBN: 978-0134610993. 6. R. Bellman: An Introduction to Artificial Intelligence: Can Computers Think? Boyd & Fraser Pub.Co., 1978. 7. R. Kurzweil: The Age of Intelligent Machines, MIT Press, 1990. 8. P.H. Winston: Artificial Intelligence, Addison-Wesley Longman Publishing, 1992. 9. N.J. Nilsson: Artificial Intelligence: A New Synthesis, Morgan Kaufmann, 1998. 10. P. Hitzler and M.K. Sarker, eds.: Neuro-Symbolic Artificial Intelligence: The State of the Art., Vol. 342, IOS Press, 2021. 11. A. Newell and H.A Simon: http://doi.acm.org/10.1145/1283920. 1283930. 12. R. Kowalski: https://doi.org/10.1017/CBO9780511984747. 13. E.A. Feigenbaum: https://doi.org/10.1016/0957-4174(92)90004-C. 14. H. Singh, et al.: https://doi.org/10.1155/2013/581879. 15. J. Pearl: Probabilistic Reasoning in Intelligent Systems - Networks of Plausible Inference, Morgan Kaufmann, 1989. 16. T.M. Mitchell: Machine Learning, McGraw-Hill, 1997. 17. L. De Raedt, et al. eds.: “Probabilistic Inductive Logic Programming - Theory and Applications,” Vol. 4911, Springer, 2008. ARTIFICIAL INTELLIGENCE APPLICATIONS IN SEMICONDUCTOR FAILURE ANALYSIS (continued from page 23)
RkJQdWJsaXNoZXIy MTMyMzg5NA==