AI-assisted circuit design with active learning

Purpose and goal

The project will investigate the benefits of "active learning" for ML models used for circuit synthetization. The models are trained with data generated via electromagnetic simulations of a large number of circuit structures. Previously, these structures have been selected in an almost completely random manner, but the amount of training data for achieving a given ML model performance could possibly be drastically reduced through smart pre-selection, which is where active learning comes in.

Expected effects and result

We expect that active learning will result in a significant reduction in the amount of training data required for a given ML model performance. The results will be presented in the form of a report and a presentation, and if they are as good as we hope, they will streamline a number of ML activities at Ericsson.

Planned approach and implementation

We will start by investigating different strategies and implementations for active learning. Then we will select the best one and compare it with the current method for data generation. Much of the work at the beginning of the thesis is about developing a special kind of ML model that is required to do smart pre-selection of new training data. Then the best model structure will be used in an active learning loop where the dataset is expanded step by step in iterations.