Embedded Intelligent Virtual Sensing (ELVIS)

Partner Call open until: October 2024

Project Start: Q1 2025

In this project we want to develop physics-based artificial intelligence (AI) algorithms to perform real-time monitoring of key parameters of electronic systems via virtual sensing. Virtual sensing uses information available from measurements and process parameters to calculate an estimate of a quantity of interest. The use of AI algorithms promises efficient computation and automatic learning of complex or even unknown relationships between available parameters and the quantity of interest.

The data from a distribution of sensors is used to monitor aspects such as:

• Temperature and mechanical stress of the system and/or the package, as these are indicators of system failure as well as factors that impact system lifetime estimation.
• Electromagnetic response of the system, to prevent interference with neighboring devices and/or to better ensure electromagnetic compliance.

The use of a physics-based AI, as opposed to more standard statistical data-driven AI algorithms, ensures a reliable operation of the virtual sensing algorithm under unexpected circumstances, since the algorithm is compliant with the laws of physics. Incorporating physics knowledge in the design of the AI model makes it less resource-hungry, which allows it to be deployed on low-resource embedded devices.

The main challenges in this project are:

• Generate suitable training data for the AI algorithm, specific for the use cases of interest and based on measurements and/or physics simulations.
• Develop physics-based neural network architectures that incorporate the essential features of the problem to ensure correctness of the predictions.
• Quantization and deployment of the AI model on embedded device, whilst preserving accuracy and real-time capabilities.
• How to combine the virtual sensing predictions with reliability models for lifetime estimation.

To tackle these challenges SAL builds upon years-long experience in the development of AI-based simulators for the thermal behavior of electronic systems, of thermal virtual sensing with neural networks as well as in implementing automated simulation workflows to generate training datasets for physics-based AI applications.

• Optimal placing of sensors for AI-based prediction of system’s temperature, mechanical stress and electromagnetic environment.
• A low-resource AI real-time virtual sensing solution deployed on a dedicated embedded device or chip.
• A thorough dataset on the physical behaviour of electronic systems, on which a multitude of other studies can be performed.

International consortium comprising 1-3 industrial partners.

• 1-3 partners
• 2-3 years runtime
• Universities and academic partners are eligible to participate (special conditions apply)
• PhD students within the project possible - supervised by academic partners

• Attractive cofinancing: SAL covers 50% of the project volume, you only pay 25% in cash and 25% in-kind.
• Competitive advantage: easier and quicker exploitation of upcoming products, by utilization of generated IP
• Minimize Risk: early start of work on technical challenges for an upcoming market
• Innovation: future technologies and exploration of emerging trends
• Partnership: cooperation in a pioneering and efficient eco-system