Barbara Hammer

KEYNOTE

Harnessing the power of deep surrogate models

Abstract: Recent advances in deep learning carry the promise to substitute computationally costly simulations or partially unobservable dynamics by deep surrogate models which are trained on example data. Since the resulting deep model is typically fast to evaluate, it is given as an explicit analytic function, and it generalizes beyond the observed training signals, deep surrogates carry diverse promises depending on the specific scenario and downstream task: They allow for a fast approximation of complex dynamic behavior; they enable real-world state inference given partial information; they support efficient system optimization based on gradient information. On the downside, deep surrogates do not only face time-consuming training of deep models, but the provision of training data is computationally complex, possibly infeasible, as it is typically based on running a complex simulation per training sample or observations of complex dynamic behavior in reality.

In the talk, I will focus on opportunities and challenges of surrogate models for a technical application, namely hydraulic simulation of water distribution systems. I will demonstrate the power of graph neural networks to learn simulations of the dynamics such that inference based on limited information becomes possible. I will have a look at how to avoid the computational challenge of huge simulations to provide training data by means of physics-informed methods, which substitute training observation by laws of physics. Further, I will have a glimpse at specific downstream tasks, which become possible based on the surrogate.