Authors: Maneesh Pawar
Abstract: The continuous scaling of semiconductor technology has significantly increased the complexity of Very Large-Scale Integration (VLSI) design. Modern integrated circuits are required to deliver high computational performance while operating within strict power and area constraints. Traditional optimization approaches based on exhaustive design-space exploration, manually tuned heuristics, and iterative Electronic Design Automation (EDA) tool execution can become computationally expensive as the number of design parameters increases. Machine Learning (ML) provides an alternative data-driven approach for predicting design quality, reducing expensive evaluations, and guiding the search toward promising design configurations. This paper presents a comparative analysis of major machine learning techniques applicable to Power, Performance, and Area (PPA) optimization in low-power VLSI design. Supervised learning methods, ensemble learning, artificial neural networks, deep learning, reinforcement learning, and graph-based learning are analyzed with respect to their suitability for PPA prediction and optimization. The study considers power consumption, timing performance, silicon area, computational cost, interpretability, scalability, training-data requirements, and generalization capability as major evaluation dimensions. A unified ML-assisted VLSI optimization framework is proposed in which circuit and synthesis features are extracted from RTL and EDA-generated reports, machine-learning models estimate PPA metrics, and an optimization engine searches for configurations satisfying multi-objective constraints. The comparative analysis indicates that no single ML technique is universally optimal for all VLSI optimization tasks. Tree-based ensemble models are attractive for structured tabular design data because of their relatively low training cost and interpretability, whereas neural and graph-based models are better suited to highly nonlinear and structural circuit relationships. Reinforcement learning is particularly promising for sequential design-space exploration and physical-design optimization, although its computational cost and dependence on reward formulation remain important limitations. The paper therefore recommends a task-oriented hybrid strategy rather than the indiscriminate use of a single ML algorithm. The proposed framework provides a reproducible foundation for future experimental validation using open VLSI benchmarks and publicly available ML-for-EDA datasets.
DOI: https://doi.org/10.5281/zenodo.23235674
