edfas.org ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 27 NO. 1 34 Flavio Cognigni discussed the role of deep learning and AI in enhancing FA capabilities, highlighting both the advancements and the challenges associated with these technologies. He explained how AI is revolutionizing tasks like automated TEM lamella preparation, cloud segmentation, 3D image reconstruction, and data denoising. However, Cognigni stresses that while AI is a powerful tool, it should not be seen as a solution for every problem. He also cautioned about the risks and ethical concerns tied to AI and deep learning, citing examples of false AI-generated results in academic papers that had passed through peer review unnoticed. After the opening statements, we continued to the discussion period. The first question from the audience spoke to those of us who wish to gain a better general understanding of machine learning and artificial intelligence. To paraphrase, the question was: “In presentations and papers on AI, we typically see diagrams containing many boxes, with one column of boxes having multiple arrows pointing to boxes in the next column and so on. Could you give us better insight into what those boxes mean?” Cognigni explained the network of boxes represents a convolutional neural network (CNN) and that the leftmost column was a list of inputs to the process, taking in the various raw characteristics of the data. The intermediate layers of columns make up the algorithms that extract and classify the features to make decisions about them. The number of intermediate layers represents the depth level of the model, and deeper models generally require more computing power to execute. The rightmost column represents the set of desired outputs from the network. Another audience member asked about the prospect of predictive analysis, where test data could be used to predict a failure mode to a reasonable degree of certainty. The panelists generally agreed that such a thing is possible, but Campbell went further and stated that his organization is doing this kind of predictive analysis currently. He gave the caveat that predictive analysis requires a good model of the failure signature, so developing a solid predictive model can be painful, but very doable. Cognigni added that his organization offers something similar to detect problems in lab tools and this technique could certainly be extended to predict failure modes in devices. The discussion shifted to the benefits and pitfalls of data sharing between organizations, particularly between design companies and foundries. Neither side wants to reveal details that might help competitors solve their own similar problems and the possibility of reverse engineering is a real risk. Campbell, Eich, and Brand spoke to the possibility of anonymizing or otherwise securing data so it could be shared between organizations, but the problem is a complex one and unlikely to have a one size fits all solution. Cognigni later mentioned that his organization was working with a major manufacturer to develop methods for anonymous data sharing. A question from the audience asked about application of AI to failures of products in low production volumes, something applicable to manufacturers in the medical and aerospace fields among others. Eich said that his organization had done some work with AI predictions for small data sets with good success. He said there are a number of ways to train machine learning models to work with small data sets, and that expert defined classification criteria and training data sets can be effective in creating viable models for assessment of small data sets. The discussion also ventured into the area of how trustworthy AI predictions really are. One question cited the errors that can be generated by chatbots such as ChatGPT. The panelists pointed out that the large language model used by chatbots is a different animal than the machine learning models that are used by AI predictive models that would be used for FA. Chatbot algorithms have far fewer constraints than any algorithm would have for analysis of electronic test data. For FA applications, the data is given a certain context and preconditioned information about what to look for. This is very different than a utility like a chatbot, which only tries to give an answer, not necessarily a correct answer. Discussion then turned to the inevitable difficulties in implementing AI-based tools into the FA process. Gap analysis and methods to predict return on investment were covered. The authors would like to thank Kim Mai of Milwaukee Tool for her meticulous transcription of the Panel Discussion. Interaction with attendees is an integral part of the User Groups.
RkJQdWJsaXNoZXIy MTYyMzk3NQ==