Macromodeling for RF/mmWave and photonics engineering
A macromodel can be defined as a mathematical model describing the behavior of a generic RF/mmWave or photonic system as seen from its inputs/outputs ports, with respect to a set of geometrical (i.e. the length and width of a component) or physical (i.e. relative permittivity and conductivity) parameters. The goal is to approximate the dynamic behavior of complex systems in terms of low-complexity models, that are efficient to evaluate. Nowadays, macromodels are applied to many design activities, ranging from design space exploration to optimization and sensitivity analysis. An ideal macromodeling technique should be able to:
- Learn complex and non-linear input/output relationships;
- Handle a large number of design variables with large parameter variations;
- Rely on a limited number of simulations/measurements (training data) to characterize the system under study
- Provide information on the level of accuracy achieved.
Unfortunately, no methodology exists complying with all these requirements, since each technique has its own advantages and drawbacks. Our research activity focuses on overcoming the current limitations of state-of-the-art macromodeling techniques by developing innovative data-efficient machine learning methods. The term data‑efficient indicates that these approaches rely on a limited amount of data (which is expensive to obtain) to reach the desired objective. Once such data‑efficient models are built, they can be integrated with machine learning approaches (for example based on Bayesian inference) to accurately and efficiently solve a large range of design problems, including optimization, design space exploration, and inverse problems.
Research topics in electrical and photonics engineering
- Surrogate-based uncertainty quantification and sensitivity analysis
- Surrogate-based design and simulation
- Surrogate-based optimization