edfas.org 23 ELECTRONIC DEVICE FAILURE ANALYSIS | VOLUME 25 NO. 2 layers are used to compute the decision, which type highly depends on the learning task. TASKS Machine learning can solve a variety of tasks as shown in Fig. 6. This section provides a short overview of the ones relevant to FA. Classification is a task when a model should assign one or more labels to a provided example. Depending on the total number of labels and the number of labels per example, one differentiates between binary, multi-class, and multilabel cases. In the first case, a classifier assigns one of two labels, in the second, one out of multiple labels, and in the third multiple out of multiple labels. Often, a classifier can output a probability for every label assignment, which can help a user to understand how confident the classifier is in its prediction. This learning task is by far the most popular one in FA applications of machine learning. For instance, in reference 37 the authors use a combined CNN-LSTM model trained to classify samples into seven failure classes with respect to the measurements of time-domain reflectometry. Zhao et al.[38] use a feedforward network to solve a binary classification problem guiding the die selection process for electrical fault isolation. The network takes a vector with 32 standard DC current measurements and dynamic current values from the various supply domains and outputs the probability for a positive physical failure analysis. Hong et al.[39] classify netlists represented as graphs into categories denoting application-specific integrated circuits. In their experiments, the authors successfully applied graph neural networks[40] to differentiate between 20 adder circuits. Application of deep learning to image classification for FA can be found in reference 41, where a CNN is used to determine the method used to create an image. The authors of reference 42 applied CNNs to identify the different logos in the printed circuit boards for the detection of counterfeit printed circuit boards. Kögel et al.[43] trained CNNs on a dataset, where each example comprised raw data output by a scanning acoustic microscope for samples that contained multiple defects in the solder bumps and labeled with a handcrafted statistical signal separation method.[44] In the test phase, their approach reached more than 97% accuracy for unseen samples. NLP approaches[45,46] classify FA reports according to their physical failures and corresponding electrical characterizations. In the first case, the authors apply classic ML techniques with TF-IDF vectorization and, in the second, deep learning methods based on a SciBERT,[47] a language representation model pre-trained on scientific texts. The latter was fine-tuned on a dataset comprising various FA documents, such as job reports or papers published in FA journals and conferences. Representation learning focuses on finding a new representation of the input data. Usually, such representation is selected to emphasize important information of the input and reduce its size and/or dimensionality. Quite often models trained for this task are used as a part of a larger machine learning system. For example, language models[45,48] are trained to represent input text as a list of vectors, which are then forwarded to the classification and clustering components, respectively. Anomaly detection is a task that aims at the detection of unusual or atypical artifacts among a large number of good artifacts. For example, in reference 37 the authors use a statistical anomaly detection method based on the mean squared distance between an observed and an ideal current measurement. Any measurement with a large distance from the normal ones is considered anomalous. Clustering is an unsupervised task often used in FA information retrieval systems to group similar artifacts, like images or texts, based on their similarity. As a result, given a new artifact, the clustering component can determine all other similar artifacts stored in the system and return them to a user. For example, in reference 48 the authors use a classical gaussian mixture model on top of a language representation to group reports of FA engineers. The object localization technique suggested in reference 49 uses specific feature extraction methods in conjunction with the K-means clustering to group regions of an image obtained with a scanning electron microscope. Then, the clusters are analyzed to determine irregular structures on an image. Regression aims to predict a numerical value for a given input. An application of regression[50] uses a random forest regressor to predict values of critical defect features, like defect size, for automated optical inspection during the ramp-up phase. OFTEN, A CLASSIFIER CAN OUTPUT A PROBABILITY FOR EVERY LABEL ASSIGNMENT, WHICH CAN HELP A USER TO UNDERSTAND HOW CONFIDENT THE CLASSIFIER IS IN ITS PREDICTION. (continued on page 26)
RkJQdWJsaXNoZXIy MTMyMzg5NA==