Deep learning techniques for integrated circuit die performance prediction
Authors
Kovalenko, A.; Lenhard, P.; Lenhard, R.
Year
2022
Published
MRS Advances. 2022, 7(30), 683-688. ISSN 2059-8521.
Type
Article
Departments
Annotation
Predicting integrated circuit (IC) functionality based on process control monitoring (PCM) parameters without individual die testing is a major challenge for manufacturers due to the high cost of electrical die measurement on the wafer. Complex dependencies between individual PCM parameters can be used to explain certain patterns of dice failure using Deep learning (DL) algorithms. However, random failure patterns due to process defects cannot be detected by this method. Combining PCM and in-process defect inspection data can be an ultimate prediction technique. In some cases, however, the availability of defect inspection data is much lower than the availability of PCM data, so direct ensemble training is rather ambiguous. This paper shows how to efficiently utilize both defect and PCM data to train a model to predict IC functionality. Such a hybrid model outperforms PCM-only solutions, and in contrast to a defect-only model predicts also failure areas across the wafer.
Integrated Circuit Die Level Yield Prediction Using Deep Learning
Authors
Lenhard, P.; Kovalenko, A.; Lenhard, R.
Year
2022
Published
2022 33rd Annual SEMI Advanced Semiconductor Manufacturing Conference (ASMC). Piscataway (New Jersey): IEEE, 2022. ISSN 1078-8743. ISBN 978-1-6654-9487-8.
Type
Proceedings paper
Departments
Annotation
Given the integrated circuits (IC) production scale, the amount of process control monitoring (PCM) data enable to develop an efficient algorithm for IC yield prediction at the die-level. Therefore, in addition to cost-effective and timeefficient yield evaluation, the proposed model is able to identify failed dice and low-yield areas on a wafer without any direct electrical die testing. Additionally, for non-parametric random dice failure detection that are untraceable by PCM input based models, an ensemble learning including both PCM and die defect inspection data are described. As Wafer Sort (WS) consumes a lot of time and resources with high associated cost a significant cost reduction can be achieved using smart product routing with selective WS by employing the aforementioned die level predictive model.
Linear Self-attention Approximation via Trainable Feedforward Kernel
Authors
Yorsh, U.; Kovalenko, A.
Year
2022
Published
Artificial Neural Networks and Machine Learning – ICANN 2022. Springer, Cham, 2022. p. 807-810. LNCS. vol. 13531. ISSN 0302-9743. ISBN 978-3-031-15933-6.
Type
Proceedings paper
Departments
Annotation
Restrictive limitation of Transformers due to the quadratic complexity of self-attention mechanism motivated a new research field of efficient Transformers, which approximate the original architecture with asymptotically faster models.
CodeDJ: Reproducible queries over large-scale software repositories
Authors
Year
2021
Published
Leibniz International Proceedings in Informatics (LIPIcs). Saarbrücken: Dagstuhl Publishing,, 2021. p. 1-24. ISSN 1868-8969. ISBN 978-3-95977-190-0.
Type
Proceedings paper
Departments
Annotation
Analyzing massive code bases is a staple of modern software engineering research – a welcome side-effect of the advent of large-scale software repositories such as GitHub. Selecting which projects one should analyze is a labor-intensive process, and a process that can lead to biased results if the selection is not representative of the population of interest. One issue faced by researchers is that the interface exposed by software repositories only allows the most basic of queries. CodeDJ is an infrastructure for querying repositories composed of a persistent datastore, constantly updated with data acquired from GitHub, and an in-memory database with a Rust query interface. CodeDJ supports reproducibility, historical queries are answered deterministically using past states of the datastore; thus researchers can reproduce published results. To illustrate the benefits of CodeDJ, we identify biases in the data of a published study and, by repeating the analysis with new data, we demonstrate that the study’s conclusions were sensitive to the choice of projects.
Dynamic Neural Diversification: Path to Computationally Sustainable Neural Networks
Authors
Year
2021
Published
Artificial Neural Networks and Machine Learning – ICANN 2021. Cham: Springer, 2021. p. 235-247. 1. vol. 12892. ISSN 1611-3349. ISBN 978-3-030-86339-5.
Type
Proceedings paper
Departments
Annotation
Small neural networks with a constrained number of trainable parameters, can be suitable resource-efficient candidates for many simple tasks, where now excessively large models are used. However, such models face several problems during the learning process, mainly due to the redundancy of the individual neurons, which results in sub-optimal accuracy or the need for additional training steps. Here, we explore the diversity of the neurons within the hidden layer during the learning process, and analyze how the diversity of the neurons affects predictions of the model. As following, we introduce several techniques to dynamically reinforce diversity between neurons during the training. These decorrelation techniques improve learning at early stages and occasionally help to overcome local minima faster. Additionally, we describe novel weight initialization method to obtain decorrelated, yet stochastic weight initialization for a fast and efficient neural network training. Decorrelated weight initialization in our case shows about 40% relative increase in test accuracy during the first 5 epochs.