AI-Based PDN Voltage Droop Prediction for Faster Configuration Evaluation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional circuit simulators are inadequate for managing complex power distribution networks (PDNs) due to challenges with multiple power rails, diverse input patterns, varying performance modes, and temperature conditions, leading to voltage deviations that impact system performance and stability.
Innovation Solution
Utilizing a convolutional recurrent neural network (CRNN) model to predict voltage overshoot and undershoot by processing time-domain current vectors and frequency-domain impedance profiles, enabling efficient evaluation and optimization of PDN configurations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If conventional circuit simulators are used to manage complex power distribution networks, then voltage deviations can be simulated, but the system becomes inefficient and time-consuming due to complexity with multiple power rails, diverse input patterns, varying performance modes, and temperature conditions
Solution Approach 1:
The patent replaces conventional circuit simulators with a deep learning-based AI model that uses convolutional neural networks and recurrent neural networks to predict voltage overshoot and undershoot. This substitution transforms the mechanical simulation process into an intelligent prediction system that can rapidly evaluate PDN configurations without time-consuming simulations, directly addressing the productivity-time loss contradiction.
Solution Approach 2:
The patent creates a virtual copy of the PDN system through AI models that learn from simulated data. The CRNN model captures the essential behavior of voltage deviations by training on synthetic PDN configurations, allowing rapid prediction without repeatedly running full circuit simulations. This copying approach enables efficient evaluation while reducing time requirements.
2Measurement precision
If conventional circuit simulators are used for PDN analysis, then voltage overshoot and undershoot can be measured, but the complexity of managing multiple power rails, diverse input patterns, and varying conditions makes the system inadequate
Solution Approach 1:
The patent extracts the essential features needed for voltage deviation prediction from complex PDN configurations. The AI model isolates key parameters such as current waveforms, impedance profiles, and temporal patterns, filtering out unnecessary complexity. This extraction enables accurate voltage overshoot and undershoot measurement while simplifying the management of multiple power rails and varying conditions.
Solution Approach 2:
The patent segments the PDN analysis into separate functional components: the CRNN model handles temporal dependencies, the CNN model processes spatial features, and intermediate models reduce impedance profiles into feature embeddings. This segmentation allows precise measurement of voltage deviations while managing complexity through modular, specialized processing stages.
3Reliability
If traditional simulation methods are used to evaluate PDN configurations, then comprehensive analysis can be performed, but the process is slow and costly
Solution Approach 1:
The patent performs preliminary training of the AI model using comprehensive simulation data to capture reliable system behavior. During training, the model learns from diverse PDN configurations under various conditions, building a foundation of reliable predictions. Once trained, the model can rapidly evaluate new configurations without requiring comprehensive re-simulation, thus maintaining reliability while improving productivity.
Solution Approach 2:
The patent implements feedback mechanisms where the AI model continuously refines its predictions based on evaluation results. The system compares predicted voltage deviations with actual performance metrics and adjusts its models accordingly. This feedback loop ensures maintained system performance stability while enabling efficient configuration evaluation through rapid iterative improvements.
Data Source
AI summary
Various embodiments include methods and computing devices implementing the methods for predicting a voltage deviation, i.e., a voltage undershoot (voltage droop) or a voltage overshoot, in a plurality of power distribution network (PDN) configurations. Various embodiments may include generating training data using a circuit simulator, in which the training data includes current waveforms and voltage waveforms associated with a plurality of PDN configurations. The generated training data may be used to train a CRNN model configured to process time-domain current vectors and frequency-domain impedance profiles. The trained CRNN model may then be used to generate a voltage deviation prediction by applying current waveforms and impedance profiles of different PDN configurations to the trained CRNN model. A plurality of PDN configuration options may be evaluated in parallel, and recommendations for PDN configurations may be determined based on the generated voltage deviation prediction and generated evaluation results.