Joshua Corsello
Joshua Corsello has a total of 3 co-authors. Here are the top co-authors by number of shared publications:
- Madhavan Swaminathan
The Pennsylvania State University
2 shared publications - Yifan Wang
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
2 shared publications - Yiliang Guo
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
2 shared publications
Latest Publications
- 12 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. PP (21563950) , 1
Deep Kernel Based Hyperparameter Adaptive Learning and Frequency Response Predictions Using Transposed Convolutional Neural Network
Electronic Design Automation (EDA) has unique challenges for addressing the design of systems for emerging applications due to the complexities involved, where multiple chiplets are integrated together on a heterogeneous platform. This challenge arises due to the long computational time required for simulation to capture all the necessary first order, second order and sometimes third order parasitic effects. Emerging machine learning (ML) and Gaussian process (GP) methods have helped expedite these processes. The deep kernel learning (DKL) model combines the structural properties of deep learning architectures with the non-parametric flexibility of kernel methods. It shows advantages by applying a GP with the corresponding kernel function to the final hidden layer of a deep neural network. However, DKL sometimes suffers from overfitting and scalability issues. In this paper, we propose an adaptive learning framework for S-parameter prediction, incorporating the spectral transposed convolutional neural network (S-TCNN) and DKL. The proposed model takes input parameters from the design space, up-samples them through transposed convolutional layers, and utilizes a GP kernel layer to approximate the desired kernel function. Additionally, the latent feature space adaptively compresses and extracts features from the input matrix, serving as a separate input parameter for the GP kernel layer. Further, we discuss the training strategy and model scalability. The proposed model is tested and evaluated using two advanced packaging examples. Results show a reduction in the number of hyperparameters by over 50% and approximately 40% improvements in loss and normalized mean-square-error (NMSE).
- 12 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. PP (21563950) , 1
Deep Kernel Based Hyperparameter Adaptive Learning and Frequency Response Predictions Using Transposed Convolutional Neural Network
Electronic Design Automation (EDA) has unique challenges for addressing the design of systems for emerging applications due to the complexities involved, where multiple chiplets are integrated together on a heterogeneous platform. This challenge arises due to the long computational time required for simulation to capture all the necessary first order, second order and sometimes third order parasitic effects. Emerging machine learning (ML) and Gaussian process (GP) methods have helped expedite these processes. The deep kernel learning (DKL) model combines the structural properties of deep learning architectures with the non-parametric flexibility of kernel methods. It shows advantages by applying a GP with the corresponding kernel function to the final hidden layer of a deep neural network. However, DKL sometimes suffers from overfitting and scalability issues. In this paper, we propose an adaptive learning framework for S-parameter prediction, incorporating the spectral transposed convolutional neural network (S-TCNN) and DKL. The proposed model takes input parameters from the design space, up-samples them through transposed convolutional layers, and utilizes a GP kernel layer to approximate the desired kernel function. Additionally, the latent feature space adaptively compresses and extracts features from the input matrix, serving as a separate input parameter for the GP kernel layer. Further, we discuss the training strategy and model scalability. The proposed model is tested and evaluated using two advanced packaging examples. Results show a reduction in the number of hyperparameters by over 50% and approximately 40% improvements in loss and normalized mean-square-error (NMSE).
- 22 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Multiscale and Multiphysics Computational Techniques
- Vol. 10, 259-270
GPU Accelerated Matrix Solution Using Novel Preconditioner for Three Dimensional Laguerre-FDTD Method
Conventionally, the large sparse matrix equation ($Ax=b$) generated by the Laguerre-FDTD method is computed using direct matrix solvers, which is often numerically expensive and computationally slow. In this work, we demonstrate an innovative approach to replace direct matrix solver with an iterative algorithm for the Laguerre-FDTD method. A novel preconditioner, specifically targeted to improve the convergence rate of biconjugate gradient stabilized solver (BiCGSTAB), is derived and implemented in the Laguerre-FDTD method. Compared with the classical Jacobi preconditioner, the proposed preconditioner achieves on average an improvement of more than 1.3× in the convergence rate. To further leverage the computational efficiency, a modified sparse matrix-vector multiplication algorithm is proposed and implemented using a General-Purpose Graphics Processing Unit (GPGPU). The new algorithm ensures that all computations are performed within the GPU, with minimum number of device-to-host data transfer and global memory access. With GPU's accelerated computing capability, the proposed solver achieves more than 5× computational speed up with respect to a high performance CPU-based direct solver on average. In addition, due to the intrinsic memory efficient nature of iterative solver, our approach also shows maximally more than 31× reduction in memory consumption against the direct solver. Various numerical examples are simulated to validate the capability and improvement of the proposed method.
- 22 May 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Multiscale and Multiphysics Computational Techniques
- Vol. 10, 259-270
GPU Accelerated Matrix Solution Using Novel Preconditioner for Three Dimensional Laguerre-FDTD Method
Conventionally, the large sparse matrix equation ($Ax=b$) generated by the Laguerre-FDTD method is computed using direct matrix solvers, which is often numerically expensive and computationally slow. In this work, we demonstrate an innovative approach to replace direct matrix solver with an iterative algorithm for the Laguerre-FDTD method. A novel preconditioner, specifically targeted to improve the convergence rate of biconjugate gradient stabilized solver (BiCGSTAB), is derived and implemented in the Laguerre-FDTD method. Compared with the classical Jacobi preconditioner, the proposed preconditioner achieves on average an improvement of more than 1.3× in the convergence rate. To further leverage the computational efficiency, a modified sparse matrix-vector multiplication algorithm is proposed and implemented using a General-Purpose Graphics Processing Unit (GPGPU). The new algorithm ensures that all computations are performed within the GPU, with minimum number of device-to-host data transfer and global memory access. With GPU's accelerated computing capability, the proposed solver achieves more than 5× computational speed up with respect to a high performance CPU-based direct solver on average. In addition, due to the intrinsic memory efficient nature of iterative solver, our approach also shows maximally more than 31× reduction in memory consumption against the direct solver. Various numerical examples are simulated to validate the capability and improvement of the proposed method.