An Explainable and Actionable Machine Learning Framework for VLSI Power and Delay Prediction
nternational Conference on Recent Progresses in Science, Engineering and Technology (ICRPSET 2026). — Third Author
Abul Basar , Abdul Kaiyum ,Mohammed Sakib Hasan ,Abdul Jobayer , Prosenijt Mondol
Abstract
In the early design stage, accurate power estimation and delay estimation is an important problem in integrated circuit (IC) design because the conventional electronic design automation (EDA) flows are time consuming and high computing cost in the synthesis and physical implementation process. While machine learning (ML) could provide a quicker solution, some methods might not be interpretable, process corner independent, or provide uncertainty quantification of the prediction. An explainable and actionable ML framework for predicting circuit power \& delay is presented. The evaluation is conducted on four regression models (XGBoost, Random Forest, LightGBM, a neural network) on a fixed train–test split. The neural network (NN) gives the best results when R2 values are target scaled and are equal to 0.905 for delay and 0.479 for power. Load capacitance and supply voltage have been found to be important features for predicting power, and transistor width, load capacitance and fan-out are significant for delay prediction using SHAP analysis. The insights are leveraged to develop a design rule for the recognition of high delay circuits with 86.4\% precision and for design correction using counterfactual analysis. Pareto-front analysis is another tool used to represent power–delay trade-offs. Cross process corner evaluation shows that the generalization is asymmetric across SS, TT and FF corners. Furthermore, the coverages of the empirical prediction interval are enhanced from 60.2\% to 89.0\% when adding the split conformal calibration. The proposed framework combines prediction, interpretability, design guidance and uncertainty estimation all to assist with making design decisions at the early stage of VLSI.
Key contributions
- 1In this research, we explore an explainable and actionable machine learning framework for predicting VLSI circuit power consumption and delay. Our work integrates multiple machine learning models, SHAP-based explainability, counterfactual design recommendations, Pareto-based optimization, and uncertainty quantification to support more reliable early-stage circuit design decisions.
Read the paper
Cite this paper
Abul Basar , Abdul Kaiyum ,Mohammed Sakib Hasan ,Abdul Jobayer , Prosenijt Mondol, "An Explainable and Actionable Machine Learning Framework for VLSI Power and Delay Prediction," in nternational Conference on Recent Progresses in Science, Engineering and Technology (ICRPSET 2026)., 2026.
Other publications
- DHT-Stack: A Deep-Hybrid Transformer Stacking Framework for Uncertainty-Aware and Cost-Sensitive Dengue Screening2026 IEEE International Conference on Biomedical Engineering, Computer and Information Technology for Health (BECITHCON)
- StrokeGuardian: Hybrid Ensemble ML for Real-Time Brain Stroke Prediction and Clinical Decision SupportIEEE WIECON-ECE 2025
- SafeRide: A Real-Time IoT-Based Smart Helmet with Road-Condition IntelligenceIEEE RAAICON 2025