Background
A digital twin is the digital model that represents a physical device, product, or even process, and is used to conduct testing, simulations, and experimentation digitally to predict the response to the physical device. Using a digital twin helps an organization reduce cost and time during design, deployment, or other phases of an enterprise product life cycle.
Implementing digital twins is extremely complex, and requires continuous model updates and training. More than 50 percent of digital twins integrate with other digital twins.1
The Metal Oxide Silicon Field Effect Transistor (MOSFET) is the most used semiconductor in various electronic products and power devices. This kit features the subthreshold voltage leakage current behavior modeling of the MOSFET digital twin. The MOSFET leakage current behavior is a key indicator for the performance of this device.
This reference kit is paired with Intel® software products, which deliver faster training, retraining, and hyperparameter tuning. By using Intel software products, organizations reduce the cost and time to value during various design and deployment product phases.
Solution
In collaboration with Accenture*, Intel developed this digital twin AI reference kit, which may assist your application to speed up training and inference of digital twin behavior modeling.
This kit includes a reference architecture for training and using an AI model with Intel® Optimization for XGBoost* to predict the leakage current.
The reference solution architecture includes using Intel Optimization for XGBoost and daal4py for training and inference optimizations to speed up inference performance. The architecture also allows for incremental learning to update the trained model.
The ideal behavior for a MOSFET is to carry zero leakage current. The digital twin model may be continuously trained and updated to predict this device behavior.
At a Glance
- Industry: Semiconductor
- Task: Predict subthreshold voltage and leakage current behavior in a MOSFET device. Iterative self-learning using pseudo-labeled data to continuously update prediction model
- Dataset: Synthetically generated leakage current
- Type of Learning: Supervised learning followed by semisupervised learning
- Models: Linear regression and XGBoost regression
- Output: Logarithm of the leakage current
- Intel® AI Software Products:
- Intel® AI Analytics Toolkit (AI Kit)
- Intel Optimization for XGBoost
- daal4py (part of Intel® Distribution for Python*)
Technology
Optimized Intel® AI Software Products for Better Performance.
AI Kit—Achieve end-to-end performance for AI workloads.
Intel Optimization for XGBoost for predictive modeling and acceleration
daal4py to accelerate inference stages
Intel has been directly upstreaming many optimizations to provide improved performance on Intel CPUs. This well-known, machine learning package for gradient-boosted decision trees now includes seamless, drop-in acceleration for Intel® architectures to significantly speed up model training and improve accuracy for better predictions.
Performance was tested on Microsoft Azure* Standard_D8_V5 using 3rd generation Intel® Xeon® processors to optimize the kit.
Benefits
Digital twins are powered by immense amounts of data to predict the outcome of a design. To be as close to a finished product as possible, data usually comes from the continuous monitoring of the real-world object, and then that data is fed back into the system. This process significantly impacts the time required to run the models at design and inference stages. A data scientist can benefit by speeding up the entire pipeline required to run a digital twin.
With this kit and Intel software products, the data scientist has the ability to train, retrain, and tune hyperparameters for digital twins faster. This may help an organization reduce overall digital twin time and cost during the overall product design and deployment phases.
Additional Resources
References
- Survey Analysis: Digital Twins Are Poised for Proliferation, Gartner document ID: G00366637, published January 31, 2019, page 3, https://www.gartner.com/document/3899678?ref=solrAll&refval=342558099
