Neural Surrogate

Neural surrogates are computationally efficient approximations of complex programs or physical simulations, primarily aimed at accelerating computations and enabling efficient Bayesian inference or optimization. Current research focuses on developing accurate and data-efficient surrogates using various neural network architectures, including coupling flows and transformers, often incorporating techniques like Bayesian meta-learning and hypernetwork-based compilation for improved performance. This field significantly impacts scientific computing by enabling faster exploration of parameter spaces, improved uncertainty quantification in complex models, and more efficient optimization of computationally expensive processes across diverse domains like atmospheric science, materials science, and cardiac simulation.