Madhavan Swaminathan
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Madhavan Swaminathan has a total of 227 co-authors. Here are the top co-authors by number of shared publications:
- Xingchen Li
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
11 shared publications - Mohanalingam Kathaperumal
D Systems Packaging Research Center, Georgia Institute of Technology, Atlanta, GA, USA
10 shared publications - Xiaofan Jia
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
8 shared publications - Ege Engin
School of Electrical and Computer Engineering, Georgia Institute of Technology, 266 Ferst Drive, Atlanta, Georgia 30332, U.S.A.
7 shared publications - Serhat Erdogan
School of Electrical and Computer Engineering, Georgia Institute of Technology, Atlanta, GA, USA
7 shared publications - Pragna Bhaskar
D Systems Packaging Research Center, Georgia Institute of Technology, Atlanta, GA 30332, USA
6 shared publications
Latest Publications
- Open Access
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (9) , 2089-2102
A Comprehensive Design Framework for Vertical Power Delivery in High-Performance Computing
- 1 September 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (9) , 2089-2102
Open AccessA Comprehensive Design Framework for Vertical Power Delivery in High-Performance Computing
- Open Access
- 13 July 2026
- Published by MDPI AG in Preprints.org
From Predict to Generate: A Cross-Layer Survey of Generative AI for IC Design and Manufacturing
AI tools used for chip design and manufacturing are moving from prediction to generation. This survey examines generative AI tools across the full silicon lifecycle. It covers analog/mixed-signal, logic and physical design, and manufacturing with yield and reliability. We organize recent work by what is generated—circuits, scripts, layouts, masks, and surrogates—and by how generation is grounded through retrieval, simulators, and physics. We give a clear taxonomy.We highlight emerging cross-stage dependencies linking early design choices to downstream manufacturability, and we identify where the literature has not yet closed those loops. We synthesize qualitative considerations for choosing retrieval-grounded LLM agents, task-specific generators, and physics-aware surrogates, while noting that controlled head-to-head evidence remains limited. We aim this work to be a broad entry point for researchers and practitioners who need a clear map and practical guidance on generative methods across the silicon lifecycle.
- 13 July 2026
- Published by MDPI AG in Preprints.org
Open AccessFrom Predict to Generate: A Cross-Layer Survey of Generative AI for IC Design and Manufacturing
AI tools used for chip design and manufacturing are moving from prediction to generation. This survey examines generative AI tools across the full silicon lifecycle. It covers analog/mixed-signal, logic and physical design, and manufacturing with yield and reliability. We organize recent work by what is generated—circuits, scripts, layouts, masks, and surrogates—and by how generation is grounded through retrieval, simulators, and physics. We give a clear taxonomy.We highlight emerging cross-stage dependencies linking early design choices to downstream manufacturability, and we identify where the literature has not yet closed those loops. We synthesize qualitative considerations for choosing retrieval-grounded LLM agents, task-specific generators, and physics-aware surrogates, while noting that controlled head-to-head evidence remains limited. We aim this work to be a broad entry point for researchers and practitioners who need a clear map and practical guidance on generative methods across the silicon lifecycle.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Multiscale and Multiphysics Computational Techniques
- Vol. 11, 291-300
Vision-Based Transformer Model for Fast Simulation of Advanced Packaging Structure
Accurate electromagnetic (EM) modeling of electronic package structures is critical for signal and power integrity analysis, yet full-wave simulations on fine meshes are often computationally expensive. Deep learning and neural network models have recently attracted considerable attention as alternative solutions. However, with the continuous expansion of the design space in advanced packaging, conventional surrogate architectures such as FCNNs, CNNs, and RNNs start to exhibit limitations in scalability. Therefore, in this paper, we propose two vision-based transformer models that eliminate the need for design parameters and leverage self-attention mechanism for efficient package-level EM and multi-physics simulation. The first framework employs a Vision Transformer (ViT) to directly predict frequency-domain responses (S-parameters) from structural images of the package. By capturing global spatial dependencies through self-attention, the ViT-based model achieves high accuracy while significantly reducing computational cost compared to conventional EM solvers. The second framework introduces a transformer-based super-resolution neural network designed for multiphysics co-simulation scenarios. Specifically, a physical structure is simulated with a coarse mesh using an EM solver to provide power input to the thermal solver, which subsequently generates low-resolution temperature profile images. The super-resolution neural network then refines these results to produce high-resolution temperature distributions, akin to those obtained through numerical simulations with a fine mesh. Various numerical examples are simulated to validate the capability and improvement of the proposed method.
- 2 June 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal on Multiscale and Multiphysics Computational Techniques
- Vol. 11, 291-300
Vision-Based Transformer Model for Fast Simulation of Advanced Packaging Structure
Accurate electromagnetic (EM) modeling of electronic package structures is critical for signal and power integrity analysis, yet full-wave simulations on fine meshes are often computationally expensive. Deep learning and neural network models have recently attracted considerable attention as alternative solutions. However, with the continuous expansion of the design space in advanced packaging, conventional surrogate architectures such as FCNNs, CNNs, and RNNs start to exhibit limitations in scalability. Therefore, in this paper, we propose two vision-based transformer models that eliminate the need for design parameters and leverage self-attention mechanism for efficient package-level EM and multi-physics simulation. The first framework employs a Vision Transformer (ViT) to directly predict frequency-domain responses (S-parameters) from structural images of the package. By capturing global spatial dependencies through self-attention, the ViT-based model achieves high accuracy while significantly reducing computational cost compared to conventional EM solvers. The second framework introduces a transformer-based super-resolution neural network designed for multiphysics co-simulation scenarios. Specifically, a physical structure is simulated with a coarse mesh using an EM solver to provide power input to the thermal solver, which subsequently generates low-resolution temperature profile images. The super-resolution neural network then refines these results to produce high-resolution temperature distributions, akin to those obtained through numerical simulations with a fine mesh. Various numerical examples are simulated to validate the capability and improvement of the proposed method.
- 31 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 44 (11) , 4600-4615
SCOPE: A Machine Learning Framework for System-Level Forward Prediction, Inverse Design, and Uncertainty Quantification in Co-Packaged Optics
This paper presents SCOPE (System-level Co-packaged Optics Prediction and Design), a machine learning framework for system-level design of co-packaged optics that models photonic integrated circuits together with the coplanar waveguide interconnects linking them to electronic integrated circuits. SCOPE operates at the link level and supports both forward performance prediction and inverse design. In SCOPE, a hybrid fully connected neural network–Bayesian neural network predicts key optical and electrical metrics from full-wave-generated data, incorporates fabrication and packaging tolerances—including misalignment as a practical factor—and provides calibrated uncertainty estimates while greatly reducing the need for full-wave and circuit co-simulations. Coupled with an actor–critic reinforcement-learning agent, SCOPE maps target specifications—such as scattering loss, and eye-diagram signal-integrity metrics—to multiple candidate photonic integrated circuit and coplanar waveguide design-parameter sets. After training, the framework identifies optimal designs whose predicted eye diagrams satisfy a specified eye mask in a few seconds per design query, yielding orders-of-magnitude speed-up compared with traditional simulation-based design flows. By capturing photonic–electronic interactions in co-packaged optics at the system level, SCOPE enables fast machine-learning-assisted design of short-reach co-packaged optical links and improves robustness to fabrication and packaging variations.
- 31 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in Journal of Lightwave Technology
- Vol. 44 (11) , 4600-4615
SCOPE: A Machine Learning Framework for System-Level Forward Prediction, Inverse Design, and Uncertainty Quantification in Co-Packaged Optics
This paper presents SCOPE (System-level Co-packaged Optics Prediction and Design), a machine learning framework for system-level design of co-packaged optics that models photonic integrated circuits together with the coplanar waveguide interconnects linking them to electronic integrated circuits. SCOPE operates at the link level and supports both forward performance prediction and inverse design. In SCOPE, a hybrid fully connected neural network–Bayesian neural network predicts key optical and electrical metrics from full-wave-generated data, incorporates fabrication and packaging tolerances—including misalignment as a practical factor—and provides calibrated uncertainty estimates while greatly reducing the need for full-wave and circuit co-simulations. Coupled with an actor–critic reinforcement-learning agent, SCOPE maps target specifications—such as scattering loss, and eye-diagram signal-integrity metrics—to multiple candidate photonic integrated circuit and coplanar waveguide design-parameter sets. After training, the framework identifies optimal designs whose predicted eye diagrams satisfy a specified eye mask in a few seconds per design query, yielding orders-of-magnitude speed-up compared with traditional simulation-based design flows. By capturing photonic–electronic interactions in co-packaged optics at the system level, SCOPE enables fast machine-learning-assisted design of short-reach co-packaged optical links and improves robustness to fabrication and packaging variations.
- 10 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (4) , 837-847
D-Band Substrate Integrated Waveguide-Fed Endfire Antenna Array in Ultrathin Glass Interposer
This article discusses the development and demonstration of a highly directive Substrate Integrated Waveguide (SIW) based Vivaldi based antenna array in package for radio applications using ultra-thin glass interposer. Beyond 100 GHz, the dimensions of the SIW in glass are small and thus are convenient to feed antennas for smart devices. The antipodal corrugations present on the arms of the Vivaldi antenna improve the radiation parameters of the proposed antenna. The SIWs are fabricated on a 2-metal layer stack-up with a 100 μm glass core (EN-A1) from AGC inc. with 15 μm of ABF GL-102 laminated on either side of the glass substrate. The semi-additive-process (SAP) which involves the formation of through-glass-vias (TGV) are also discussed. The use of glass substrates enables high repeatability of the fabricated antenna due to process control. The article presents a procedure for normalized 3D radiation pattern of the fabricated antennas. The simulated and measured results are presented for the normalized radiation pattern, boresight gain, and return loss of the antenna array in D-band. The measured results show that the array achieved a maximum gain of 12.9 dBi and 14.5 dBi at boresight for a 4-element and an 8-element array respectively.
- 10 March 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (4) , 837-847
D-Band Substrate Integrated Waveguide-Fed Endfire Antenna Array in Ultrathin Glass Interposer
This article discusses the development and demonstration of a highly directive Substrate Integrated Waveguide (SIW) based Vivaldi based antenna array in package for radio applications using ultra-thin glass interposer. Beyond 100 GHz, the dimensions of the SIW in glass are small and thus are convenient to feed antennas for smart devices. The antipodal corrugations present on the arms of the Vivaldi antenna improve the radiation parameters of the proposed antenna. The SIWs are fabricated on a 2-metal layer stack-up with a 100 μm glass core (EN-A1) from AGC inc. with 15 μm of ABF GL-102 laminated on either side of the glass substrate. The semi-additive-process (SAP) which involves the formation of through-glass-vias (TGV) are also discussed. The use of glass substrates enables high repeatability of the fabricated antenna due to process control. The article presents a procedure for normalized 3D radiation pattern of the fabricated antennas. The simulated and measured results are presented for the normalized radiation pattern, boresight gain, and return loss of the antenna array in D-band. The measured results show that the array achieved a maximum gain of 12.9 dBi and 14.5 dBi at boresight for a 4-element and an 8-element array respectively.
- 25 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (4) , 693-702
Embedded High-Power RF Chips in Glass Core Package
Chip-embedded glass packaging provides a promising solution for RF system integration. However, its performance at higher power levels hasn’t been documented. This paper presents the electrical and thermal performance of an RF power amplifier (PA) embedded in a glass package. A 22.5-W GaN-PA embedded glass package has been designed, fabricated and tested. The results show that the deviation of the electrical performance between the bare die and packaged PA is minimal. In addition, the packaged PA showed good thermal performance at a power density of 0.89 W/mm2. Also, the model and measurements show good correlation.
- 25 February 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. 16 (4) , 693-702
Embedded High-Power RF Chips in Glass Core Package
Chip-embedded glass packaging provides a promising solution for RF system integration. However, its performance at higher power levels hasn’t been documented. This paper presents the electrical and thermal performance of an RF power amplifier (PA) embedded in a glass package. A 22.5-W GaN-PA embedded glass package has been designed, fabricated and tested. The results show that the deviation of the electrical performance between the bare die and packaged PA is minimal. In addition, the packaged PA showed good thermal performance at a power density of 0.89 W/mm2. Also, the model and measurements show good correlation.
- 24 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. PP (21563950) , 1
Self-Consistent Electrothermal Modeling of Distributed Vertical Power Delivery Architecture With Substrate-Embedded Microfluidic Cooling
Lateral power delivery (LPD) becomes infeasible for high-performance computing (HPC) processors operating at voltages near 1 V with current densities of 2 A·mm−2 due to large routing losses and Joule heating. Distributed vertical power delivery (DVPD) with integrated voltage regulators (IVRs) addresses this challenge by placing conversion stages close to the load, while substrate-embedded microfluidic cooling provides effective heat-dissipation pathways for inner tiers in vertically dense stacks. However, frameworks for self-consistent electrothermal modeling of DVPD at the system scale remain underexplored. This work develops such a framework, achieving high physical fidelity and computational efficiency. These attributes are realized by enforcing geometric and power profile consistency between the electrical and thermal models and through automation and optimized meshing. Applied to a 48-to-1 V DVPD architecture delivering 1 kW to a 500 mm2 die at 2 A·mm−2, the framework yields a self-consistent solution near the 85 °C threshold with a total system loss of 231 W and a maximum temperature of 85.7 °C, requiring a mass flow rate of 1.39 g·s−1 in embedded microchannels. Compared to an uncoupled initial estimate of 200 W with an initial mass flow rate of 1 g·s−1, the total system loss increases by about 15.8 %, which shows that neglecting electrothermal coupling underestimates both losses and cooling requirements. Additionally, region-specific meshing balances mesh resolution by refining near critical features and coarsening elsewhere, reducing both meshing and per-iteration times by over 30.0 % without loss of prediction accuracy.
- 24 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Components, Packaging and Manufacturing Technology
- Vol. PP (21563950) , 1
Self-Consistent Electrothermal Modeling of Distributed Vertical Power Delivery Architecture With Substrate-Embedded Microfluidic Cooling
Lateral power delivery (LPD) becomes infeasible for high-performance computing (HPC) processors operating at voltages near 1 V with current densities of 2 A·mm−2 due to large routing losses and Joule heating. Distributed vertical power delivery (DVPD) with integrated voltage regulators (IVRs) addresses this challenge by placing conversion stages close to the load, while substrate-embedded microfluidic cooling provides effective heat-dissipation pathways for inner tiers in vertically dense stacks. However, frameworks for self-consistent electrothermal modeling of DVPD at the system scale remain underexplored. This work develops such a framework, achieving high physical fidelity and computational efficiency. These attributes are realized by enforcing geometric and power profile consistency between the electrical and thermal models and through automation and optimized meshing. Applied to a 48-to-1 V DVPD architecture delivering 1 kW to a 500 mm2 die at 2 A·mm−2, the framework yields a self-consistent solution near the 85 °C threshold with a total system loss of 231 W and a maximum temperature of 85.7 °C, requiring a mass flow rate of 1.39 g·s−1 in embedded microchannels. Compared to an uncoupled initial estimate of 200 W with an initial mass flow rate of 1 g·s−1, the total system loss increases by about 15.8 %, which shows that neglecting electrothermal coupling underestimates both losses and cooling requirements. Additionally, region-specific meshing balances mesh resolution by refining near critical features and coarsening elsewhere, reducing both meshing and per-iteration times by over 30.0 % without loss of prediction accuracy.
- 7 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Microwave Theory and Techniques
- Vol. 74 (1) , 63-72
Integrated Circuit-on-Glass (ICoG): A Self-Packaged 3D-Heterogeneous Integration (3DHI) Platform for Millimeter-Wave Circuits With Embedded GaN-on-Si Dielets
This work proposes a novel integrated circuit-on-glass (ICoG) platform for the 3D-heterogeneous integration (3DHI) of highly scaled gallium nitride (GaN) dielets for fabrication of millimeter-wave (mmWave) circuits. Au-free GaN-on-Si high electron mobility transistor (HEMT) dielets are fabricated using a copper gate technology and singulated to a size of $350\times 540$ $\mu $ m. The GaN dielets are embedded into a glass interposer featuring a two-layer redistribution layer (RDL) and copper back-end-of-line (BEOL) fabricated using semi-additive (SAP) processing. After quantifying the RDL loading effects on the GaN, a model of the dielet after integration is developed. This model was then used for the design and fabrication of two amplifiers, one targeting the fifth-generation (5G) new radio (NR) frequency range 2 (FR2) 28 GHz band and a 10 GHz amplifier featuring high linearity for future 5G NR frequency range 3(FR3) applications.
- 7 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Microwave Theory and Techniques
- Vol. 74 (1) , 63-72
Integrated Circuit-on-Glass (ICoG): A Self-Packaged 3D-Heterogeneous Integration (3DHI) Platform for Millimeter-Wave Circuits With Embedded GaN-on-Si Dielets
This work proposes a novel integrated circuit-on-glass (ICoG) platform for the 3D-heterogeneous integration (3DHI) of highly scaled gallium nitride (GaN) dielets for fabrication of millimeter-wave (mmWave) circuits. Au-free GaN-on-Si high electron mobility transistor (HEMT) dielets are fabricated using a copper gate technology and singulated to a size of $350\times 540$ $\mu $ m. The GaN dielets are embedded into a glass interposer featuring a two-layer redistribution layer (RDL) and copper back-end-of-line (BEOL) fabricated using semi-additive (SAP) processing. After quantifying the RDL loading effects on the GaN, a model of the dielet after integration is developed. This model was then used for the design and fabrication of two amplifiers, one targeting the fifth-generation (5G) new radio (NR) frequency range 2 (FR2) 28 GHz band and a 10 GHz amplifier featuring high linearity for future 5G NR frequency range 3(FR3) applications.