Addressing A Posteriori Performance Degradation in Neural Network Subgrid Stress Models

A recent study highlights the challenges of estimating prediction uncertainty in active learning campaigns for materials discovery, indicating that uncalibrated out-of-distribution uncertainty quantification may hinder model performance. The research evaluates various uncertainty estimation methods using ensembles of ALIGNN, eXtreme Gradient Boost, Random Forest, and Neural Network architectures, focusing on tasks related to solubility, bandgap, and formation energy predictions.

The FAIR-Pruner method has been introduced to enhance neural network pruning by adaptively determining the sparsity levels of each layer, addressing the limitations of traditional uniform pruning strategies that often lead to performance degradation. This innovative approach utilizes a novel indicator, Tolerance of Differences (ToD), to balance importance scores from different perspectives, thus improving efficiency in resource-limited environments.

A new approach called local entropy search (LES) has been proposed for Bayesian optimization, focusing on refining design spaces through gradient descent. This method propagates posterior beliefs over objectives, resulting in a probability distribution that guides the selection of the next evaluation by maximizing mutual information. Empirical results indicate that LES demonstrates strong sample efficiency compared to existing optimization methods.