ORNL advances hyperspectral AI processing with D-CHAG method

Oak Ridge National Laboratory has developed D-CHAG, a new method that more than doubles processing speed and reduces memory usage by up to 75% for hyperspectral plant imaging, enabling training of larger AI foundation models on the Frontier exascale supercomputer.

• Main announcement: ORNL researchers introduced Distributed Cross-Channel Hierarchical Aggregation (D-CHAG), which uses distributed tokenization across GPUs and hierarchical aggregation to split and merge spectral channels; demonstrated >2x processing speed and up to 75% memory reduction on APPL hyperspectral data and a weather dataset, enabling training of larger foundation models without loss of spatial or spectral resolution. Presented at SC25 (Nov 2025) and run on Frontier at the Oak Ridge Leadership Computing Facility.
• Background and details: The work supports DOE’s Genesis Mission and DOE-supported projects including OPAL and a Generative Pretrained Transformer for Genomic Photosynthesis; collaborators and supporters include the Center for Bioenergy Innovation, ORNL laboratory-directed research & development funding, and partner institutions such as Argonne, Lawrence Berkeley, PNNL, and RIKEN; next steps include refining models to predict photosynthetic efficiency directly from hyperspectral images.