COURSE 2: Data-driven surrogates in multiscale modeling
Course description:
| Day and Time | Monday 17th of February 2025, all day, within 9 am and 5 pm (exact schedule to be confirmed) |
| Place | Arts et Métiers – ENSAM (Paris Campus) |
| Total teaching hours | Full-day course |
| Attendance Fee | 250 euros |
| Maximum Attendance | 90 Participants |
One of the potential breakthroughs that machine learning offers in computational mechanics is that data-driven surrogates may enable two-way coupled multiscale analysis. There is a widely shared vision to replace the expensive micromodels that need to be evaluated many times in computational homogenization (or FE2) with surrogates trained on a data set from a smaller number of micromodel evaluations. However, realizing this vision requires that the surrogate consistently gives accurate predictions on unseen scenarios and that the surrogate response is sufficiently smooth for convergence of the macromodel. Well-informed choices need to be made in designing the architecture of the surrogate, in model selection, and in training and validation. This course is aimed at young researchers with a background in computational mechanics. Participants will be provided with a solid foundation for performing data-driven multiscale analysis. By attending the course, they will be equipped to apply recently proposed surrogate modeling strategies to new problems, or to develop and test new surrogate modeling strategies.
Objectives and target groups:
In this one-day course, we start with machine learning fundamentals to reinforce the understand- ing of popular tools (k-nearest neighbor models, neural networks, Gaussian processes). Different approaches for constructing data-driven surrogates in the context of multiscale modeling will be introduced with applications from recent literature. In an interactive workshop, hands-on exercises will be given based on python implementations of several surrogate approaches:
- Feed-forward neural network
- Recurrent neural network (GRU)
- Hybrid architecture with physics-based ingredients
Course Outline:
| Time | Topic | Instructor(s) |
|---|---|---|
| 09:00-10:30 | Machine learning fundamentals | Rocha | 10:30-11:00 | Coffee Break |
| 11:00-12:30 | Overview of surrogate modeling strategies | Van der Meer | 12:30-13:30 | Lunch Break |
| 13:30-15:00 | Workshop | Van der Meer, Rocha, Maia | 15:00-15:30 | Coffee Break |
| 15:30-17:00 | Workshop, Conclusion and open challenges | Van der Meer |
Preliminary Time Schedule
The course will be a full-day course consisting of two 90-minute slots.
Lecturers:
