SBOCF Optimizes Image-Based Inverse Problems in Materials.

A new method called Scalable Bayesian Optimization of Composite Functions (SBOCF) has been developed to address inverse problems in materials characterization, specifically estimating physical parameters from scientific images. These problems often rely on computationally expensive physics-based simulations. For instance, in electron microscopy, determining specimen thickness and crystal mistilt is crucial for accurate atomic-scale structure recovery, typically done by matching experimental patterns with simulated ones. Traditional grid searches are inefficient, and neural network approaches demand extensive pretraining that may not transfer well to new conditions. SBOCF is designed to be simulation-efficient by leveraging the known composite structure of the image-matching objective and the intermediate information available in simulated images. It represents complex PACBED images with patch-level summaries and correction terms, effectively reducing the number of modeled outputs from over 24,000 to just 11 while preserving the original pixel-wise objective. In synthetic benchmarks, SBOCF significantly outperformed standard Bayesian optimization, achieving up to a 290x reduction in median final SSE for thick samples within a budget of 50 simulator evaluations. When applied to experimental data, SBOCF produced parameter estimates consistent with established values and improved downstream ptychographic reconstruction, leading to sharper atomic images.