Machine learning based reduced-order models to predict spatiotemporal dynamics of soil carbon and biomass yield of different bioenergy crops

Abstract

Agroecosystem models are commonly used to predict the effects of management practices and environmental changes on biomass yields, and soil organic carbon (SOC) dynamics under different bioenergy crops. However, the computational and data requirements of these models limit their scalability across multiple scenarios and timescales, particularly when trying to capture the range of outcomes under different scenarios. Machine learning (ML) offers a way to streamline these complex processes, enabling efficient modeling with substantially less computational resources. In this study, we combined ML model with an agroecosystem model Daily Century (DAYCENT) to project baseline (2009–2018) and future (2021–2100) biomass yields and SOC changes for three bioenergy crops:Miscanthus, sorghum, and switchgrassacross U.S. agricultural lands. These projections were made under the Shared Socio-economic Pathway 8.5 (SSP5 8.5), using data from the coupled model intercomparison project phase six models. The ML-based reduced order model, trained on field observations and DAYCENT outputs, accurately predicted baseline biomass yields with R² values ranging from 0.96 to 0.98, and SOC changes with R² values between 0.93 and 0.98 across the three bioenergy crops. Under a SSP5 8.5 scenario, Miscanthus and sorghum exhibited lower sensitivity to precipitation and temperature impacts in terms of biomass yield and SOC changes compared to switchgrass. Sorghum and Miscanthus are projected to see an increase in economically viable land area across the continental U.S., with gains of 29 % and 10 %, respectively. The most significant increases for sorghum are expected at higher latitudes. In contrast, economically viable land area is projected to decline by 11 % by 2100 compared to baseline scenarios for switchgrass. Our findings demonstrate the potential of ML-based reduced order models to provide accurate predictions offering opportunity to develop user-friendly agroecosystem analysis tools in future.