3rd Webinar NumPEx – PDE Neural Surrogates

A PDE neural surrogate is a neural network trained to approximate a PDE solver, with the promise of delivering a solution at a fraction of the cost of a classical numerical simulation. This talk will survey the main existing approaches, from physics-informed neural networks (PINNs) to data-driven surrogates trained in a supervised manner on solver-generated datasets. We will then examine the challenges that such approaches still need to overcome, from generalization capabilities to compute costs, including both data generation and training costs. Finally, we will focus on online training, where data generation and model training are performed concurrently. We will present the solution provided by the Melissa framework, which enables large-scale online surrogate training with up to 10× acceleration compared to classical file-based training.

By Bruno Raffin