The following projects are currently being conducted within the Graduate School. Click on a project's title to learn more about the project and the involved investigators.
Information about open positions in the Graduate School can be found under this address.
Prof. Dr. Jens Anders
Institut für Intelligente Sensorik und
Prof. Dr. Ilia Polian
Institut für Technische Informatik
Abt. Hardwareorientierte Informatik
Co-supervisor: Dr. Matthias Sauer, Advantest
Test quality, defined as the absence of test escapes (defective circuits that had passed post-manufacturing test), is the ultimate target of testing. Customers apply system-level test (SLT) to circuits that already have been tested post-fabrication and reportedly identify test escapes. The objective of this project is to understand the nature of such hard-to-detect failures. Establishing a better understanding of SLT and making it more effective and efficient could drastically improve the economy of circuit design and manufacturing.
A number of theories exist for the type of failures that cause SLT-unique fails that are missed by post-manufacturing tests:
- Complex defect mechanisms with subtle parametric or intermittent manifestations that are not adequately covered by standard fault models. The test coverage of such defects can be improved by the use of more advanced defect-oriented models.
- Systematic ATPG coverage holes: Insufficient coverage of structures such as clock domain boundaries, asynchronous or analog interfaces, clock distribution networks and other sources of unknown values. Standard automatic test pattern generation (ATPG) tools tend to classify faults in such structures as “untestable” even though they can manifest themselves during normal operation of the device.
- Marginal defects exposed only during system-level interactions: Subtle defects, in particular related to timing, in “uncore logic” of complex multicore systems on chip (SoCs), i.e., logic that is part of the SoC but does not belong to a core.
The specific objectives of the project are as follows:
- To establish a theoretical and systematic understanding of SLT-unique fails, identifying specific mechanisms leading to such fails and their manifestation conditions (e.g., hot-spots).
- To create an experimentation environment where SLT-unique fails can be reproduced and practically investigated.
- To explore solutions that prevent or address SLT-unique fails. These can include guidelines for “clean” circuit designs that do not give rise to coverage holes (e.g., use of well-defined asynchronous protocols for clock domain crossings); design for testability (DFT) methods that address specific weaknesses known to the designer or DFT engineer; extended ATPG methods that can detect defects missed by regular ATPG, or methods to create effective and efficient SLTs that specifically target SLT-unique failure mechanisms.
The project is structured into three tasks according to the three above-mentioned scientific objectives. The more theoretical Tasks 1 and 3 deal with SLT-unique fails and solutions to counteract them, respectively. Task 2 will establish a complete evaluation and experimentation flow that can be used for practical demonstration of SLT-unique fails and studies of applicable solutions. Figure 1 summarizes the planned project structure and the interaction of its theoretical (red) and practical (blue) parts. An SLT evaluation platform ① will be created and SLT-unique fail conditions ② from Task 1 will be incorporated into this platform. Based on the outcome of experiments ③, solutions for the SLT-unique fails ④ from Task 3 (e.g., addition of special DFT logic) will be incorporated into the SLT evaluation platform ①, thus closing the loop.
Prof. Dr.-Ing. Bin Yang
Institute of Signal Processing
and System Theory
Co-supervisors: Jochen Rivoir and Raphael Latty, Advantest
Post-silicon validation deals with the test of devices under test (DUT) in order to find and correct design bugs before mass production. For doing this, up to several hundreds of input variables or features are recorded. They characterize the input stimuli to the DUT, various tuning parameters and environmental conditions. At the same time, some target variables are calculated from the responses of the DUT. By studying the relationship between the input and target variables, design bugs and unexpected effects have to be detected, localized and mitigated. Today this is still done manually by experienced engineers. However, several hundreds of input variables are too much for visualization and manual inspection. Since a single target variable is typically related to a few input variables, the selection of relevant input variables for a specific target variable becomes a crucial problem.
Numerous traditional methods have been developed for the task of feature or variable selection, e.g. wrapper/filter/embedded methods. They were successful in some applications and failed in others. The goal of this project is to take a fresh look at the old problem of variable selection from a new perspective of deep learning. Deep learning in this context does not mean a deep and large neural network as a black box for everything. We rather mean an increasing number of recent and successful ideas and architectures developed for deep learning. Some of them have a strong relationship to the problem of variable selection. We aim to adopt these ideas to variable selection and to develop new approaches which hopefully outperform the traditional methods. Of course, also a combination of the traditional methods and new approaches is highly desirable.
Two first ideas of deep learning based variable selection to be studied are:
- Attention-based variable selection. An attention network is used to compute an attention vector containing weights to select or deselect the individual input variables. Here, the selection is considered as a binary classification problem (select or deselect) and the corresponding weights denote the probabilities of selection. The weighted input vector is the input for a second evaluation network which evaluates the ability of the weighted input vector to predict the desired target variable.
- Concrete autoencoder (AE). AE is a well-known architecture to learn a nonlinear hidden lower-dimensional representation for the given input. This technique is a nonlinear extension of the conventional principal component analysis (PCA). The bridge between AE and variable selection is the use of a Concrete distribution, a continuous relaxation of discrete random variables. By doing this, the continuous-valued weight matrix of the encoder can be trained by normal backpropagation. During inference, each column of the weight matrix approaches a one-hot vector for a zero temperature limit and thus selects a single feature.
This project is closely related to two other projects: P2 “Visual analytics for post-silicon validation” and P3 “Self-learning test case generation for post-silicon validation”. P2 uses the results from this project to visualize the relationship between the target variable and the selected input variables. P3 employs self-learning methods to generate more test cases which will improve the variable selection. On the other hand, the results of variable selection will provide an improved understanding of the input-target-relationship to guide the self-learning test case generation.