print · login   

Physics-informed neural networks in microelectronics environment (internship at Infineon)

Are you interested in applying Machine Learning Algorithms in Microelectronics Design? Are you passionate about combining knowledge and concepts from Mathematics, Physics, Neural Networks, Electromagnetic Simulation and Measurements? We are currently looking for a student who would push the boundaries of our knowledge of Neural Networks and their application in Modelling.

Description of the project:

The focus of the project will be on the physics-informed neural networks (PINNs), and their applicability in real-world applications at Infineon PINNs represent a cutting-edge integration of traditional physics-based modelling and machine learning techniques. These innovative models leverage neural networks to learn and approximate physical phenomena while incorporating fundamental principles of physics to enhance accuracy and reliability in predictions. In Infineon, we want to identify the possibility of injecting Physics Differential Equations of interest in the Neural Networks training algorithm. We would like to see the capabilities of such an approach in a microelectronics environment where we need to model a parameter that needs to respect the laws of physics. When we use mathematical models, sometimes, we can encounter numerical aberrations that do not allow the simulation to converge.

Additionally, we can use this method to model the degradation of a transistor over time; the path that a transistor follows is the solution of a complex differential equation.

We would like to make a trial where we use this method to solve complex differential equations and calculate multidimensional integrations. We will start with simple single-dimensional equations and, in the course of the internship, try to scale up the method to higher dimensions.

Exploiting PINNs allows us to resolve one of the biggest critiques of Neural Networks, namely that they cannot extrapolate well beyond the data used for the training. One way to go overcome this limitation is to train the ANN differently and make it respect a certain relationship between the integral, derivative and the output itself.

About Infineon

Infineon designs, develops, manufactures, and markets a broad range of semiconductors and semiconductor-based solutions, focusing on key markets in the automotive, industrial, and consumer sectors. Its products range from standard components to special components for digital, analogue, and mixed-signal applications to customer-specific solutions together with the appropriate software.

Contact: Yuliya Shapovalova (RU).