Andi Buzo

Andi Buzo

Auto-generated profile, based on publicly available information

Andi Buzo

Infineon Technologies AG Neubiberg, Neubiberg, Germany

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
30 January 2025 in AEU - International Journal of Electronics and Communications
Verified
Journal Article
Open access
EPEleni Papageorgiou
Eleni Papageorgiou
ABAndi Buzo
Andi Buzo
GPGeorg Pelz
Georg Pelz
Thomas NoulisThomas Noulis
Thomas Noulis

18 September 2024 in IEEE Access
Verified
Journal Article
Open access
ARAlecsandra Rusu
Alecsandra Rusu
EDEmilian David
Emilian David
VGVasile Grosu
Vasile Grosu
MTMarina Topa
Marina Topa
ABAndi Buzo
Andi Buzo
GPGeorg Pelz
Georg Pelz
This paper presents an adaptive pre-silicon integrated circuit verification algorithm that incorporates a machine learning algorithm. This approach can overcome the traditional corner-based process-voltage-temperature verification limitations considering variables coverage. The proposed solution employs an adaptive algorithm whose parameters have been analyzed in detail and optimized to achieve high performance. The optimal configurations for these parameters were demonstrated through the accuracy and efficiency of the algorithm. Additionally, this paper presents three methods that expand the use of the univariate algorithm to multivariate adaptive scenarios. These techniques aim to achieve the most accurate identification of worst-case circuit behavior through simultaneously modeling multiple electrical parameters (EP) outputs. The effectiveness of the proposed methods was validated through extensive testing on a large and diverse set of synthetic test functions that intend to replicate the behavior of real circuits. The algorithms consistency and accuracy are also validated on real Low Dropout Voltage Regulator (LDO) circuits.

9 October 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
EPEleni Papageorgiou
Eleni Papageorgiou
VDValentina Davidoiu
Valentina Davidoiu
GPGeorg Pelz
Georg Pelz
ABAndi Buzo
Andi Buzo
Thomas NoulisThomas Noulis
Thomas Noulis
Conventional reinforcement learning (RL) frameworks for analog sizing frequently suffer from arbitrary action-space discretization, initialization bias from localized starting points, and sample inefficiency under strict multi-corner constraints. This paper presents a target-conditioned RL framework featuring sensitivity-guided multi-step action spaces and curriculum learning for robust, end-to-end analog sizing. Parameter adjustment steps are analytically calibrated through normalized sensitivity analysis to balance coarse global exploration with fine local refinement. To alleviate sparse rewards under tight performance bounds, a progressive curriculum strategy tightens specification intervals across stages, augmented by a 'lazy' evaluation scheme that filters unpromising designs at the nominal corner to drastically cut simulation overhead. Furthermore, a distributed, antidiametric multi-point initialization benchmark is introduced to quantitatively diagnose policy neutrality and solution dispersion across the design space. Evaluated on a two-stage operational amplifier across 27 PVT corners, the proposed method achieves a 70.0% specification satisfaction rate, outperforming uncalibrated baselines by up to 7x while reducing simulation cost by over 85%, and demonstrates unbiased convergence across diverse starting regions.

26 August 2026 in SSRN Electronic Journal
Verified
Preprint
Open access
SKSavvas Karipidis
Savvas Karipidis
VDValentina Davidoiu
Valentina Davidoiu
ABAndi Buzo
Andi Buzo
GPGeorg Pelz
Georg Pelz
Thomas NoulisThomas Noulis
Thomas Noulis
A novel generative deep learning surrogate modeling framework is proposed for predicting the full continuous time-domain dynamic transient response of non-linear analog circuits, thus eliminating the large signal analysis from the custom design flow. The circuit vehicle used is a 12-parameter cold-start Dickson frequency-boosting charge pump in a 180 nm CMOS technology. To overcome the exponential simulation burden of high-dimensional grid sampling and the limitations of simplified closed-form equations, an automated data generation pipeline is introduced. The pipeline integrates dual-decoupled scrambled Sobol sequence sampling, SipHash duplicate filtering, and parallelized Ramer-Douglas-Peucker (RDP) waveform compression. The core surrogate model employs a hybrid deep architecture consisting of a fully connected feature-expansion stage and an 1D transposed convolutional decoder stack. Automated Bayesian hyperparameter optimization using a tree-structured Parzen estimator (TPE) identifies the optimal network structure and training dynamics. The surrogate model synthesizes full continuous transient waveforms with a mean normalized dynamic voltage error of 10.30 % (standard deviation of 4.75 %) and an inference execution time of 25 ms, providing high-fidelity transient prediction and massive design cycle acceleration.