Machine Learning for Simulation

So far, classical simulations and Machine Learning (ML) are separated fields based on different paradigms. This Project Network paves the way for integrating them with joint model-based and data-driven predictions capable of tackling simulation problems that are currently out of reach. In addition, opening the black box of ML through visualization will give non-experts access to these tools. On the one hand, the network will focus methodologically on supervised, unsupervised, and reinforcement learning. We will develop, analyze, and implement novel ML strategies adapted to the characteristics of simulation data, such as symmetries, invariances, and physical prior knowledge. And, surprisingly, even deterministic trajectories are incompatible with statistical assumptions required for standard learning guarantees for ML methods.

RQ 1 ML4Sim: How can ML be adopted for simulation; specifically, how can we incorporate application-specific physical constraints and prior knowledge, and how can we learn in the absence of ML-specific statistical assumptions?