News Updates
Press releases related to the CROSSROADS-SSA can be found here.
Congratulations to Dominik! September 2026.
Dominik Bittner won the best oral presentation award at the 5th International Soil Modelling Consortium Conference in Rio da Janeiro. Dominik's talk was on: 'Differentiable hybrid model integrating RothC and neural networks to predict soil organic carbon dynamics'. Abstract is below and more information on the conference can be found here.
Abstract:
Soil is essential for providing ecosystem services, including food production and climate regulation. Soil organic carbon (SOC) is key to this, as it is the largest active terrestrial carbon reservoir and is vital to the global carbon cycle. Therefore, predicting SOC dynamics in a changing climate is crucial for sustainable land management. The dynamics of SOC turnover are commonly simulated using multi-compartment, process-based models of decomposition, such as RothC, which offer mechanistic interpretability and explicit process representation. However, these models face some limitations, including simplified representation of processes, a dependence on steady-state assumptions for initialisation and high computational costs at large spatial scale. With growing data availability, artificial intelligence (AI) approaches are increasingly used in soil mapping. Models, such as long short-term memories and recurrent neural networks, are used to predict time-series data, such as SOC dynamics, as they can capture complex, non-linear patterns. However, purely data-driven models typically lack physical consistency, perform poorly in extrapolation tasks and might not adhere to physical laws. The limitations of both types of models are being addressed in ongoing research that links process-based and AI models by feeding the output from one model into another. However, this can propagate errors, and the results remain difficult to interpret. Here we present a differentiable hybrid modeling framework that uses RothC as a backbone, with neural networks embedded to refine site-specific parameters in a physically consistent manner. Leveraging automatic differentiation to obtain gradients of model outputs with respect to inputs, the framework enables efficient calibration of SOC pools and carbon inputs, avoiding the computational burden of traditional spin-up approaches. We apply this new methodology to data from long-term experimental sites in temperate croplands where SOC is measured repeatedly. We compare results to the traditional standalone process-based and purely data-driven approaches. We also analyse the capacity of this new approach to derive intermediate response functions. The results suggest our approach could bridge the gap between process-based and AI-driven methods, facilitating the development of scalable, interpretable and physically consistent simulations, and enabling the identification of previously under-represented patterns and processes in SOC dynamics.
Professorship awarded to Awdenegest Moges. Link ]
Honorary professorship recognises Ethiopian collaborative research. Link ]
Better ways to communicate impacts of under-utilised plant species in the Bilate catchment , Ethiopia Newsletter
