EDA² | ISEDA 2025 | Technical Session

Emerging Transistor Technologies and Novel Process Innovations

Abstract: Accurate extraction of light intensity and wavelength from photodetector responses is essential for decoding information, yet the intertwined influence of the two parameters on device performance poses significant challenges. This study presents a neural network-based model that extracts light intensity and wavelength directly from the current-voltage (I-V) characteristics of 2D materials-based photodetectors. Using a shared-bottom multi-task learning framework, a one-dimensional convolutional neural network (1D-CNN) was employed to capture feature representations from the I-V data, enabling precise classification and extraction of incident light parameters. The model was trained and evaluated using a dataset comprising 3750 I-V curves measured from a high-dynamic range and broadband WSe2/MoS2 heterojunction photodetector. The model exhibits an overall classification accuracy of 99.7%, highlighting the deep learning approach as a data-driven alternative to conventional analytical methods for modeling 2D photodetectors.

Abstract: In semiconductor manufacturing, the deposition quality of thin film plays a critical role in both production efficiency and device performance. Plasma-enhanced Chemical Vapor Deposition (PECVD) has emerged as a widely adopted technique due to its low process temperature and high deposition rate. A comprehensive understanding of the thin film deposition mechanism of PECVD is essential for optimizing process parameters. This study introduces a Spatiotemporal Attention-enhanced ConvLSTM (STAE-ConvLSTM) model for predicting thin film deposition, specifically modeling the SiOxNy deposition process in PECVD. By integrating advanced spatiotemporal prediction and self-attention techniques, the model effectively captures global spatiotemporal dependencies and the influence of process parameters on film growth. Through extensive testing on diverse datasets, the proposed model demonstrates superior performance in predicting thin film deposition compared to both standard ConvLSTM networks and traditional physical-chemistry-based models. The experimental results highlight the STAE-ConvLSTM model’s ability to handle complex substrate structures with high precision. Additionally, long-sequence prediction experiments confirm the model’s robustness. The proposed model can predict thin film deposition process effectively.

Abstract: In this work, a hard-constraint (HC) enhanced Physics-Informed Neural Network (PINN) framework for accurate device-level thermal simulations is developed, effectively addressing challenges such as complex boundary conditions, multi-material systems, and interface continuity constraints. The approach integrates the domain decomposition method and the mixed residual method (MIM) while enforcing boundary and interface conditions through a hard-constraint strategy, ensuring an improved simulation accuracy. The proposed enhanced PINN thermal simulation framework is verified through its application to advanced silicon-on-insulator (SOI) MOSFETs. Compared with conventional PINN and extended physics-informed neural networks (XPINN) methods, the approach achieves a 4.76× improvement in the MAPE of temperature predictions and significantly enhances the heat flux continuity at material interfaces, ensuring better adherence to physical laws. This work provides a scalable and high-fidelity solution for complex device-level thermal simulations and future semiconductor thermal evaluation and optimization.

Abstract: This study proposes an automated extraction method for HEMT device modeling from technical documents using AI - agent - based models. A data stream - based automation method is proposed, enabling direct HEMT model generation from datasheets. The system includes document parsing, graphical data extraction, parameter extraction using a Chain - of - Thought (CoT) prompting strategy, and SPICE modeling. The ASM - HEMT model, a physics - based compact model, accurately describes GaN HEMTs' properties. The proposed data stream employs EDocDLA for document parsing, EDocCurve for curve extraction, and integrates LLMs and optimization techniques for parameter extraction and model optimization. Experiments on the FHX76LP datasheet demonstrate the data stream's effectiveness. The generated SPICE model has a fitting error below 3% compared to the original data. This automated method significantly enhances modeling efficiency and accuracy, providing an efficient pathway for transforming technical documents into precise device models.