- Laura Mainini, Politecnico Di Torino
- Matteo Diez, CNR-INM, National Research Council-Institute of Marine Engineering
Machine learning is becoming very popular to compute approximations of physical processes and emulators (digital twins and siblings) of engineering systems from measurements and black box simulations. However, commonly used machine learning techniques (e.g. neural networks) require significant amounts of data to learn from; in addition, the responses are frequently affected by lack of interpretability and their reliability is of difficult characterization. These features constitute major limitations to the use of these techniques for engineering applications associated with time-critical and safety-critical decisions, for which the collection of reference data points is usually expensive. Examples of problems of this kind include the assessment and optimization of design alternatives, reliability assessment, health diagnostics and prognostics of systems and structures. The scientific community is dedicating efforts to address those limitations and propose advanced formulations that combine scientific computing and machine learning. Potential approaches leverage the synergies between machine learning and model reduction methods, multifidelity modelling, uncertainty characterization. The objective of this minisymposium is to bring together researchers with different backgrounds to discuss the insights of different computational methods and their applications to engineering problems.