Preprint: Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training
Latest Melissa related paper available on Arxiv: [2606.09949] Learning Where to Simulate: Generative Active Sampling for Online PDE Surrogate Training
Melissa training workflow is extended to support active learning and automatically run simulations generating more relevant data for training. Adiffusion neural network is trained in parallel with the surrogate to infer the parametrization of the next set of simulations to run to generate “hard-to-learn” samples. Results with different PDEs, neural models, in 1D and 2D, ifering up to 300 parameters, show significant gains on the tail statistics. See paper for details.
