Cheng Shi, Maruti Mudunuru, Maggie Bowman, Qian Zhao, Jason Toyoda, William Kew, Yuri Corilo, Odeta Qafoku, John R Bargar, Satish Karra, Emily Graham

Despite the importance of microbial respiration of soil organic matter (SOM) in regulating carbon flux between soils and the atmosphere, soil carbon (C) cycling models remain primarily based on climate and soil properties, leading to large uncertainty in their predictions. Molecular data have long been proposed as a promising avenue for resolving modeling errors, but evidence for improved predictions of soil C cycles with high-resolution measurements remains mixed. With data from the 1000 Soils Pilot of the Molecular Observation Network (MONet), we developed a workflow to analyze the molecular composition of water-extractable SOM from 66 soil cores across the United States to address this knowledge gap. Our innovation lies in using machine learning (ML) to distill the thousands of SOM formula that we detected per sample into tractable units; and it enables data from state-of-science measurement techniques to be filtered into the molecules that most directly explain soil respiration. Then, we compared ML predictions of measured potential soil respiration using (1) a suite of standard soil physicochemical data, (2) ultrahigh-resolution SOM composition independently, and (3) in combination with physicochemistry to assess the added value of molecular information to predict soil respiration. In surface soils (0-10 cm), SOM chemistry alone provided better estimates of potential soil respiration than soil physicochemical factors alone, and using the combined sets of predictors yielded the best prediction of soil respiration. In contrast, in subsoils (>10 cm), SOM composition did not improve respiration model performance, possibly due to the importance of mineral-associated SOM below the surface layer. Our workflow is applicable to multiple types of mass spectrometry data and to studies on environmental changes ranging from localized experiments to global surveys. We underscore the advances of ML tools in downscaling the thousands of SOM molecules detected by state-of-science mass spectrometry for developing new carbon cycling models.