Physics-inspired ML framework for Compact Model Calibration

Recent advances in machine learning (ML) have introduced a promising alternative path for CM calibration. ML-based methods automate parameter extraction by learning complex, non-linear relationships between measurements and model parameters, reducing manual effort and accelerating calibration. Techniques such as neural networks, Bayesian optimization, and surrogate modelling handle high-dimensional parameter spaces and improve robustness against noise and variability, while transfer learning aids adaptation to new devices. Despite these advantages, ML approaches face challenges including the need for large, high-quality datasets, risk of overfitting, and integration with SPICE and EDA workflows to maintain physical consistency. Interpretability also remains a concern, as black-box models can obscure the physical meaning of parameters, making validation and trust critical for industrial adoption.

                                   Internship/Master thesis Scope

This internship will focus on developing physics-inspired ML methods for calibrating CMs. Unlike purely data-driven approaches, physics-inspired ML integrates fundamental device principles—such as charge transport, electrostatics, and temperature dependence—into the ML architecture or loss function. This hybrid approach combines the predictive power of ML with physical interpretability, reducing overfitting and improving generalization to unseen bias or temperature conditions. By leveraging prior knowledge, these methods can work effectively with smaller datasets, making them particularly suitable for emerging technologies where measurement data is limited. Key challenges include designing architectures that balance flexibility with physical constraints and ensuring seamless integration into SPICE-compatible workflows for circuit simulation. The ultimate goal of the internship is to create a physics-informed ML framework for CM calibration that complements existing in-house calibration tools.