Search

Compact models for CMOS transistors usually have many fitting parameters to accurately capture the device properties, especially for the cutting-edge CMOS technology. As a result, parameter extraction of compact models requires a lot of expertise and engineering time. To overcome this barrier, we propose a new machine learning approach, graph-based compact model (GCM), to automate parameter extraction with high efficiency. GCM starts from a core set of physical equations, such as long-channel surface potential with semi-empirical analytic equations for short-channel effects. It then aggregates these physical models through graph neural networks (GNNs) to predict the final device behavior. In this approach, the analytic equations preserve physical dependencies on process and bias conditions, while the neural networks in GCM enable model training driven by a small set of measurement data. Using GCM, we demonstrate parameter extraction with high accuracy for dc and ac data from GlobalFoundries 12 nm FinFET technology. We further incorporate channel length and temperature dependence in GCM. The generation of a full GCM model card is less than 5 min, all automated through the back propagation process. Finally, GCM is implemented in Verilog-A and passes Si2 benchmark tests, ensuring model continuity and quality in circuit simulations.

In this paper, we propose a graph model embedded with compact physical equations for modeling the thermal dynamics of buildings. The principles of heat flow across various components in the building, such as walls and doors, fit the message-passing strategy used by Graph Neural networks (GNNs). The proposed method is to represent the multi-zone building as a graph, in which only zones are considered as nodes, and any heat flow between zones is modeled as an edge based on prior knowledge of the building structure. Furthermore, the thermal dynamics of these components are described by compact models in the graph. GNNs are further employed to train model parameters from collected data. During model training, our proposed method enforces physical constraints (e.g., zone sizes and connections) on model parameters and propagates the penalty in the loss function of GNN. Such constraints are essential to ensure model robustness and interpretability. We evaluate the effectiveness of the proposed modeling approach on a realistic dataset with multiple zones. The results demonstrate a satisfactory accuracy in the prediction of multi-zone temperature. Moreover, we illustrate that the new model can reliably learn hidden physical parameters with incomplete data.

ABSTRACTBackground: We investigated the influence of miR-144 on the cisplatin-sensitivity of anaplastic thyroid carcinoma (ATC) cells and explored the internal molecular mechanism of miR-144.Methods: Thyroid cancer cells ARO, TPC1 and normal thyroid cells HT-ori3 were used in this research. Expressions of miR-144 and TGF-α were uncovered by western blot and qRT-PCR. Expressions of autophagy-related protein LC3 II and apoptosis-related protein Caspase-3 and PARP were explored by western blot and immunofluorescence. Cell viability was detected by MTT assay and apoptosis condition was revealed by flow cytometric analysis and TUNEL staining. Dual-luciferase reporter assay was employed to verify the target relationship. Tissue sections were detected by IHC. Xenograft assay was conducted to further verify conclusions in vivo.Results: MiR-144, which was low expressed in ATC cells and tissues, could inhibit autophagy activation induced by cisplatin, enhancing the sensitivity of ATC cells to cisplatin, and promoting cell apoptosis. TGF-α was the target of miR-144 and was negatively regulated by it. MiR-144 could improve the sensitivity of ATC cells to cisplatin and inhibit tumor growth by suppressing TGF-α both in vitroand in vivo.Conclusion: MiR-144 could inhibit autophagy of ATC cells by down-regulating TGF-α, enhancing the cisplatin-sensitivity of ATC cells.