Omer Rochman

Hey! I'm PhD student at ULiège, under the supervision of Prof. Gilles Louppe. My main interest are neural emulation and solutions of XDEs, and I'm currently exploring the potential of diffusion models to solve forward and inverse PDE problems. Previously, I worked as an AI researcher at appliedAI, a Munich based AI consultancy. I did a master's degree in Mathematics at the Technical University of Munich (TUM) and a bachelor's degree in Physics at the University of Leipzig.

Email  /  CV  /  Twitter  /  BlueSky  /  Github

We studied different projections methods to enforce governing equation constraints (i.e. residuals) after training, and the effects the enforcements have on the trajectories of PDEs.

Appa is a large weather model (1.5B) composed of an autoencoder and a latent diffusion model. After unsupervised training, it can assimilate any kind of non-linear observations to generate plausible trajectories, in tasks such as reanalysis or forecasting. We demonstrate its strong and promising results and flexibility.

NeuralMPM can emulate point-cloud fluid systems that include multiple materials, each with their specific properties, at over 1000 FPS, vs 15 for the original simulator. It only needs trajectories of positions and velocities to learn, bypassing the tricky tuning of a simulator.

We evaluated different stabilization strategies in the context of graph-based neural emulators (GNS).

We propose DiffComp, a symbolic regressor that can learn arbitrary function compositions, including derivatives of various orders.

We attempted to solve the Time Dependent Schrödinger Equation using PINNs, finding that they are unable to, and examine that failure.