Compact models of transistors are essential for simulating and optimizing circuits with the use of SPICE simulation tool. Parameter extraction, which is calibrating these models, is essential to ensure their alignment with measured or simulated data. However, conventional parameter extraction methods are generally iterative and experience-dependent, requiring significant time and effort from modeling engineers. Moreover, as semiconductor devices and compact models become increasingly advanced, the need for a tailored extraction process for each model has become increasingly inefficient. To address the above challenges, this work proposes an agile transistor compact modeling framework, ATMAD. The proposed framework takes a compact model file and a set of electrical characteristic data as inputs, producing a calibrated model with minimal human intervention. ATMAD automatically retrieves the equations in the compact model and converts them into computational flow graphs, thus supporting different compact models with a generalized process. A graph unlooping technique is proposed to support automatic differentiation for compact models with implicit functions (e.g., series resistance and surface potential solving). Based on the computational flow graph, ATMAD adopts automatic differentiation technique to achieve automatic and parallel optimization of model parameters. The proposed ATMAD framework is validated on commonly-used compact models in academia and industry, showing its effectiveness for compact modeling for both I-V and C-V characteristics.

Advanced semiconductor lithography faces significant challenges as feature sizes continue to shrink, necessitating effective Multiple Patterning Layout Decomposition (MPLD) algorithms. Existing MPLD algorithms are inefficient or cannot support stitch insertion to achieve finer-grained optimal decom-position. This paper introduces an end-to-end GNN-based frame-work that not only achieves high-quality solutions quickly but also applies to layouts with stitches. Our framework treats layouts as heterogeneous graphs and performs inference through a message-passing mechanism. We deliver ultra-competitive, near-optimal solutions that are 10x faster than the exact algorithm (e.g., integer linear programming) and 3x faster than approximate algorithms (e.g., exact-cover, semi-definite programming).

Healthcare Education and Labeling for Cardiopulmonary Sounds, HEALSound, is an interactive platform designed to enhance the identification and labeling of adventitious cardiopulmonary sounds through gamification. HEALSound enables users, particularly medical professionals, to engage in exercises that improve their knowledge of abnormal heart and lung sounds while simultaneously contributing to the labeling of raw audio data. By incorporating real-time feedback and progress tracking, the app promotes continuous learning and provides a valuable crowdsourced resource for building high-quality labeled datasets over time. This dual-purpose platform not only aids in medical education but also contributes to the advancement of machine learning models for sound classification in healthcare. The system demonstrates a new potential for significant impact in educational and clinical environments by seamlessly integrating learning with data collection, ensuring scalability and the continuous improvement of cardiopulmonary sound databases.

Advances in digital stethoscopes and wearable health sensors now support real-time cardiopulmonary monitoring and AI-driven diagnostic tools. However, the extensive manual labeling required for large datasets of respiratory sounds presents a significant barrier, traditionally dependent on expert input. To address this, we introduce HEALSound, a mobile application designed to blend educational training with crowdsourced data labeling. By engaging users in interactive auscultation exercises and providing immediate feedback, HEALSound promotes skill development while generating high-quality labeled data through weighted consensus methods. Experimental results highlight the app’s dual effectiveness: enhancing diagnostic learning for users and accelerating the development of machine-learning models through enriched datasets, thus addressing critical challenges in AI-based health diagnostics.

Modeling lithography using machine learning is extremely data-intensive. Due to intellectual property privacy concerns and potential malicious attacks, design houses and foundries are unwilling to share their designs directly. To address the aforementioned concerns, we have proposed a personalized federated learning-based framework (pFed-Litho) to perform end-to-end lithography simulation. This framework incorporates a cross-level local training algorithm along with an integrated optimization method to generate personalized and local models, which overcome the generalization problem and slow convergence and oscillatory behavior in its loss function, respectively. The experimental results show that our pFed-Litho framework achieves up to 14.07% higher accuracy with reduced oscillatory behavior in the loss curve compared to the state-of-the-art works. Even with a dataset reduced by$100\times $, our framework maintains a stable accuracy of over 91%, representing a 50% increase compared to the U-Net model.

This work introduces an innovative wearable health-care system that offers personalized cardiopulmonary monitoring by continuously capturing a variety of physiological signals. Users can engage with the system to view their own data in real-time, thereby gaining a nuanced understanding of its multi-modal sensing capabilities and the potential for remote health monitoring.

Continuous physiological sound monitoring is crucial for the prevention, diagnosis, and treatment of various diseases like cardiopulmonary and gastrointestinal conditions. Wearable healthcare sensors have emerged as a potent solution, streamlining the capture, storage, transmission, and analysis of individualized physiological sounds. However, challenges exist including large data volumes, limited hardware computational capabilities, and constrained transmission bit rates. To address these issues, we propose a physiological sound compression system using compressive sensing with self-adaptive compression ratio across sound types to implement physiological sound compression and Optimized Discrete Cosine Transform (ODCT) reconstruction to reduce loss in effective bands. Evaluated on SPRSound and PhysioNet 2016, our approach attains correlation coefficients of 0.863 and 0.883 for respiratory and cardiac sounds, with -3.14 dB and -1.84 dB signal-to-noise ratio loss at 3.5 and 3.0 compression ratios. Implemented on a custom healthcare sensor, our approach optimizes bit rate to 1.73× and power consumption to 0.82× compared to the uncompressed system.

Malicious manipulation of very large-scale integration physical-layout design is a serious problem in modern integrated circuit design. The physical-layout design database requires a highly compressed secured storage medium. In this article, we propose a secured compressive asymmetrical convolutional auto-encoder (ACAE) machine learning framework, CompressKey, which performs layout compression and encryption simultaneously. It utilizes geometric features to eliminate redundancies in layout patterns. We propose a “Divide and Merge” technique to partition each layer into smaller sizes of unique patterns to address the inconsistency of layout pattern complexity. We also propose “Matrix Expansion” and “Matrix Reduction” techniques on the matrix-based pattern to achieve secured “near lossless” compression on the layouts. We have evaluated CompressKey on 14/28/32 nm open-source ICCAD contest databases and achieved a secured compression ratio of 4.54 with encryption features outperforming$1.22\times $–$1.59\times $compared to the state-of-the-art techniques.

Performing chemical mechanical polishing (CMP) modeling for physical verification on an integrated circuit (IC) chip is vital to minimize its manufacturing yield loss. Traditional CMP models calculate post-CMP topography height of the IC’s layout based on physical principles and empirical experiments, which is computationally costly and time-consuming. In this work, we propose a CmpCNN framework based on convolutional neural networks (CNNs) with a transfer learning method to accelerate the CMP modeling process. It utilizes a multi-input strategy by feeding the binary image of layout and its density into our CNN-based model to extract features more efficiently. The transfer learning method is adopted to different CMP process parameters and different categories of circuits to further improve its prediction accuracy and convergence speed. Experimental results show that our CmpCNN framework achieves a competitive root mean square error ( RMSE ) of 2.7733Å with 1.89× reduction compared to the prior work, and a 57× speedup compared to the commercial CMP simulation tool.

With Moore's law reaching its limits, the use of new materials or new devices' structure has emerged as the next generation of CMOS devices. Among all, tunneling field-effect transistors (TFETs) have achieved a steep sub-threshold slope of less than 60mV/decade yet there is a lack of a complete process design kit (PDK) for large-scale circuit design. Hence, we present an open-source 15nm TFET PDK (15nmTFETPDK), which is based on FreePDK15 with additional support for open-source TFET models and its associated Cadence Virtuoso pcell in OA format and Mentor Calibre physical verification files. Our users can design their circuit in the Cadence Virtuoso platform and verify the consistency of the drawn layout and the circuit through the Calibre Platform. We hope that this open-source PDK allows them to further advance their research with new emerging devices.

The use of deep neural networks in pattern matching has tremendously improved the accuracy of the lithographic hotspot detection, preventing any catastrophic chip failure. In this paper, we propose three data augmentation techniques ("Translation", "Gaussian noise", and "Fill shapes") to deal with the imbalance outlier lithographic hotspot problem and adopt the neural ordinary differential equations networks (Litho-NeuralODE) to improve the detection accuracy. Our architecture uses 28x28 pixel clips to perform the hotspot classification. Experimental result on ICCAD 2012 Contest benchmarks shows that our proposed framework achieves the overall highest accuracy of 98.7% and the lowest misses of 10 on average, outperforming the state-of-the-art works.