Search

Abstract This study demonstrates a novel use of the U-Net convolutional neural network (CNN) for modeling pixel-based electrostatic potential distributions in GaN metal–insulator-semiconductor high-electron mobility transistors (MIS-HEMTs) with various gate and source field plate designs and drain voltages. The pixel-based images of the potential distribution are successfully modeled from the developed U-Net CNN with an error of less than 1% error relative to a TCAD simulated reference of a 500-V electrostatic potential distribution in the AlGaN/GaN interface. Furthermore, the modeling time of potential distributions by U-Net takes about 80 ms. Therefore, the U-Net CNN is a promising approach to efficiently model the pixel-based distributions characteristics in GaN power devices.

The brain atlas, which provides information about the distribution of genes, proteins, neurons, or anatomical regions, plays a crucial role in contemporary neuroscience research. To analyze the spatial distribution of those substances based on images from different brain samples, we often need to warp and register individual brain images to a standard brain template. However, the process of warping and registration may lead to spatial errors, thereby severely reducing the accuracy of the analysis. To address this issue, we develop an automated method for segmenting neuropils in the Drosophila brain for fluorescence images from the FlyCircuit database. This technique allows future brain atlas studies to be conducted accurately at the individual level without warping and aligning to a standard brain template. Our method, LYNSU (Locating by YOLO and Segmenting by U-Net), consists of two stages. In the first stage, we use the YOLOv7 model to quickly locate neuropils and rapidly extract small-scale 3D images as input for the second stage model. This stage achieves a 99.4% accuracy rate in neuropil localization. In the second stage, we employ the 3D U-Net model to segment neuropils. LYNSU can achieve high accuracy in segmentation using a small training set consisting of images from merely 16 brains. We demonstrate LYNSU on six distinct neuropils or structures, achieving a high segmentation accuracy comparable to professional manual annotations with a 3D Intersection-over-Union (IoU) reaching up to 0.869. Our method takes only about 7 s to segment a neuropil while achieving a similar level of performance as the human annotators. To demonstrate a use case of LYNSU, we applied it to all female Drosophila brains from the FlyCircuit database to investigate the asymmetry of the mushroom bodies (MBs), the learning center of fruit flies. We used LYNSU to segment bilateral MBs and compare the volumes between left and right for each individual. Notably, of 8,703 valid brain samples, 10.14% showed bilateral volume differences that exceeded 10%. The study demonstrated the potential of the proposed method in high-throughput anatomical analysis and connectomics construction of the Drosophila brain.

In this work, the methodology using Graph Attention Network (GAT) for the reserve design in GaN power MIS-HEMTs based on hand-drawn characteristics is demonstrated for the first-time. The hand-drawn ID-VG characteristic is constructed by Ramer-Douglas-Peucker algorithm. Then, the extracted information is sent to the Graph Attention Network to receive the corresponding device design variables, including tAlGaN, recessed depth, Al%, Lg, Lgd, and Lgs. Less than 30 seconds is consumed to generate the design variables and less than 8% of the differences in the key extracted parameters, such as threshold voltage (Vth), On-state current (Ion), and subthreshold slope (SS), can be achieved by comparing hand-drawn ID-VG and simulated ID-VG characteristic based on the design variables from GAT model. Therefore, the developed GAT approach is promising for the reverse design of GaN power MIS-HEMTs, which can provide users with efficient and valuable design suggestions to optimize the devices toward the targeting performance.

Abstract Although Technology Computer-Aided Design (TCAD) simulation has paved a successful and efficient way to significantly reduce the cost of experiments under the device design, it still encounters many challenges as the semiconductor industry goes through rapid development in recent years, i.e. Complex 3D device structures, power devices. Recently, although machine learning has been proposed to enable the simulation acceleration and inverse‑design of devices, which can quickly and accurately predict device performance, up to now physical quantities (such as electric field, potential energy, quantum-mechanically confined carrier distributions, and so on) being essential for understanding device physics can still only be obtained by traditional time-consuming self-consistent calculation. In this work, we employ a modified U-Net and train the models to predict the physical quantities of a MOSFET in two-dimensional landscapes for the first time. Errors in predictions by the two models have been analyzed, which shows the importance of a sufficient amount of data to prediction accuracy. The computation time for one landscape prediction with high accuracy by our well-trained U-Net model is much faster than the traditional approach. This work paves the way for interpretable predictions of device simulations based on convolutional neural networks.

Abstract Three-dimensional integrated circuit (3D IC) technologies have been receiving much attention recently due to the near-ending of Moore’s law of minimization in 2D IC. However, the reliability of 3D IC, which is greatly influenced by voids and failure in interconnects during the fabrication processes, typically requires slow testing and relies on human’s judgement. Thus, the growing demand for 3D IC has generated considerable attention on the importance of reliability analysis and failure prediction. This research conducts 3D X-ray tomographic images combining with AI deep learning based on a convolutional neural network (CNN) for non-destructive analysis of solder interconnects. By training the AI machine using a reliable database of collected images, the AI can quickly detect and predict the interconnect operational faults of solder joints with an accuracy of up to 89.9% based on non-destructive 3D X-ray tomographic images. The important features which determine the “Good” or “Failure” condition for a reflowed microbump, such as area loss percentage at the middle cross-section, are also revealed.

Abstract Engineering simulation accelerates the development of reliable and repeatable design processes in various domains. However, the computing resource consumption is dramatically raised in the whole development processes. Making the most of these simulation data becomes more and more important in modern industrial product design. In the present study, we proposed a workflow comprised of a series of machine learning algorithms (mainly deep neuron networks) to be an alternative to the numerical simulation. We have applied the workflow to the field of dental implant design process. The process is based on a complex, time-dependent, multi-physical biomechanical theory, known as mechano-regulatory method. It has been used to evaluate the performance of dental implants and to assess the tissue recovery after the oral surgery procedures. We provided a deep learning network (DLN) with calibrated simulation data that came from different simulation conditions with experimental verification. The DLN achieves nearly exact result of simulated bone healing history around implants. The correlation of the predicted essential physical properties of surrounding bones (e.g. strain and fluid velocity) and performance indexes of implants (e.g. bone area and bone-implant contact) were greater than 0.980 and 0.947, respectively. The testing AUC values for the classification of each tissue phenotype were ranging from 0.90 to 0.99. The DLN reduced hours of simulation time to seconds. Moreover, our DLN is explainable via Deep Taylor decomposition, suggesting that the transverse fluid velocity, upper and lower parts of dental implants are the keys that influence bone healing and the distribution of tissue phenotypes the most. Many examples of commercial dental implants with designs which follow these design strategies can be found. This work demonstrates that DLN with proper network design is capable to replace complex, time-dependent, multi-physical models/theories, as well as to reveal the underlying features without prior professional knowledge.

Tissue differentiation varies based on patients’ conditions, such as occlusal force and bone properties. Thus, the design of the implants needs to take these conditions into account to improve osseointegration. However, the efficiency of the design procedure is typically not satisfactory and needs to be significantly improved. Thus, a deep learning network (DLN) is proposed in this study. A data-driven DLN consisting of U-net, ANN, and random forest models was implemented. It serves as a surrogate for finite element analysis and the mechano-regulation algorithm. The datasets include the history of tissue differentiation throughout 35 days with various levels of occlusal force and bone properties. The accuracy of day-by-day tissue differentiation prediction in the testing dataset was 82%, and the AUC value of the five tissue phenotypes (fibrous tissue, cartilage, immature bone, mature bone, and resorption) was above 0.86, showing a high prediction accuracy. The proposed DLN model showed the robustness for surrogating the complex, time-dependent calculations. The results can serve as a design guideline for dental implants.

Segmenting individual neurons from a large number of noisy raw images is the first step in building a comprehensive map of neuron-to-neuron connections for predicting information flow in the brain. Thousands of fluorescence-labeled brain neurons have been imaged. However, mapping a complete connectome remains challenging because imaged neurons are often entangled and manual segmentation of a large population of single neurons is laborious and prone to bias. In this study, we report an automatic algorithm, NeuroRetriever, for unbiased large-scale segmentation of confocal fluorescence images of single neurons in the adult Drosophila brain. NeuroRetriever uses a high-dynamic-range thresholding method to segment three-dimensional morphology of single neurons based on branch-specific structural features. Applying NeuroRetriever to automatically segment single neurons in 22,037 raw brain images, we successfully retrieved 28,125 individual neurons validated by human segmentation. Thus, automated NeuroRetriever will greatly accelerate 3D reconstruction of the single neurons for constructing the complete connectomes.

Abstract High resolution synchrotron microtomography capable of revealing microvessels in three dimensional (3D) establishes distinct imaging markers of mouse kidney disease strongly associated to renal tubulointerstitial (TI) lesions and glomerulopathy. Two complementary mouse models of chronic kidney disease (CKD), unilateral ureteral obstruction (UUO) and focal segmental glomerulosclerosis (FSGS), were used and five candidates of unique 3D imaging markers were identified. Our characterization to differentially reflect the altered microvasculature of renal TI lesions and/or glomerulopathy demonstrated these image features can be used to differentiate the disease status and the possible cause therefore qualified as image markers. These 3D imaging markers were further correlated with the histopathology and renal microvessel-based molecular study using antibodies against vascular endothelial cells (CD31), the connective tissue growth factor or the vascular endothelial growth factor. We also found that these 3D imaging markers individually characterize the development of renal TI lesions or glomerulopathy, quantitative and integrated use of all of them provide more information for differentiating the two renal conditions. Our findings thus establish a practical strategy to characterize the CKD-associated renal injuries by the microangiography-based 3D imaging and highlight the impact of dysfunctional microvasculature as a whole on the pathogenesis of the renal lesions.

A device design technique using tunneling barriers (TBs) for reducing the short-channel effects (SCEs) is proposed. By introducing TBs at the source and drain junctions of a Si FET, the threshold voltage (Vth) roll-off can be significantly suppressed. This is because the TBs weaken the electrical coupling between drain bias and transmission/current spectrum in energy. Specifically, as compared with a conventional FET, the Vth roll-off for channel length reduction from 20 to 5 nm is mitigated by more than 40% when a thin TB is embedded at the source junction. This paper further reveals that the TB at the source junction dominates the physical mechanism minimizing the SCEs of the TBFET, and thus the device performance can be improved appreciably by removing the TB at the drain side and by decreasing the TB height at the source side. [ABSTRACT FROM AUTHOR]

An abstract of the article "Large-scale segmentation and tracing for neurons in drosophila brain by Fast Automatically Structural Tracing Algorithm (FASTA)" by Nan-Yow Chen and colleagues is presented.