Conception and application of machine learning for inductance prediction in multilayer rectangular spiral micro coils

Introduction/purpose: This research introduces a novel approach for designing a dataset of multilayer rectangular planar coils by the integration of a complementary software tools. MATLAB works as a high-level design environment, facilitating the creation of complex geometries and FastHenry acts as a computational engine to solve Maxwell's equations and extract inductance values. Two diverse synthetic datasets are generated using advanced sampling techniques, including Latin hypercube sampling, for different configurations. These datasets are then processed and trained using machine learning algorithms to predict inductance values based on the derived geometric parameters. Methods: Initially, MATLAB is used to generate extensive synthetic datasets, comprising 20000 rows for 2-layer coil configurations and 15000 rows for 3-layer configurations. After the generation process, the datasets are checked for the readiness for training. Six machine learning models ( Gaussian Process Regressor (GPR), KNeighborsRegressor (KNN), BayesianRidge, ElasticNetCV, GammaRegressor, and Bagging Regressor) are trained and evaluated using metrics such as R² and RMSE. The models are further tested on unseen test data and validated using the cross-validation technique to check how much the models can generalize. Results: The datasets were generated successfully, and the models KNeighborsRegressor, Gaussian Process Regressor (GPR), and Bagging Regressor performed the best and showed a high accuracy and low error. Conclusion: The results show that machine learning is a practical and effective method for predicting inductance in multilayer rectangular planar coils based on the geometry.