Electrical measurements in electronic, electrochemical, and biomedical systems rarely provide direct access to quantities of interest. Equivalent circuit models (ECMs) address this gap by linking measured responses to underlying system properties through idealized circuit elements. However, ECM identification from electrical measurements is an ill-posed inverse problem, and model ambiguity and fitting errors may propagate to subsequent tasks. Although data-driven machine learning (ML), as an alternative interpretation tool, can learn complex nonlinear mappings from measured data without explicit models, it often lacks interpretability and physical consistency.

This thesis addresses these challenges by developing physics-informed learning frameworks that integrate ECM-based structural information with ML across three progressively deeper levels. First, ML enhances ECM fitting via a deep learning regression model trained on ECM-simulated data to pre-fit ECM parameters from electrical impedance spectroscopy (EIS). This method facilitates in-vivo bioimpedance analysis of bone fractures. Second, a physics-informed neural network surrogate constrained by circuit laws is developed for rapid diode-based circuit performance evaluation from device voltage–current characteristics. Third, a multi-task framework embeds a differentiable simulator within a neural network, leveraging ECM parameter extraction as an auxiliary task to improve battery state estimation from EIS with physical interpretability.

Overall, this thesis demonstrates the mutual reinforcement between parameterized circuit models and data-driven ML, highlighting the potential of ECM-ML integration as a generalizable framework for interpreting electrical measurements.

After the defence there will be a small reception at Fredrik Bajers Vej 7, A4-106