While the functions carried out by most of the vital organs in humans are unremarkable, the human brain clearly separates us from the rest of life on the planet. It’s a vastly complicated tissue and little is known about its cellular microstructure; particularly its synaptic circuits are almost completely unexplored. These circuits underlie the unparalleled capabilities of the human mind, and when disrupted, likely underlie some of the incurable disorders of brain function. 

Three advances make it possible to pursue a precise understanding of the structure of the brain today: next-generation electron microscopes are capable of imaging with multiple beams simultaneously to speed the enormous task of imaging tissue at high resolution; accelerator-based computing is pushing supercomputers to exascale and beyond; and large deep learning models are quickly outperforming humans at laborious tasks such as identifying neurons in imaged data. This INCITE proposal aims to leverage all of these to produce datasets of human brain connectivity at unprecedented scale, for analysis within a separately funded neuroscience-driven project, and to publish the data via ALCF’s Globus- based data sharing facilities. 

While we acknowledge that reconstruction of a whole human brain with eighty billion neurons is a task for microscopes and supercomputers one or two generations beyond those available today, our work today necessarily pushes the boundaries toward achieving those future goals. This extreme-scale, AI-driven project extends this team’s earlier Aurora Early Science project to adapt their neural network-based segmentation code for Polaris and Aurora. Already team has found that trained models have the ability to generalize across datasets, though fine-tuning is of course required; amid the ocean of data processed in this INCITE project, the team hopes this work enables steps towards a foundation model for segmentation, capable of superior performance across future EM datasets with minimal fine- tuning.