The physically-based process model like STEMMUS-SCOPE can serve perfectly as the virtual laboratory to study responses of ecosystem functioning to various climate stressors (e.g., rising CO2, temperature, and increasingly frequent extreme drought events). However, a major bottleneck using such advanced model, in routine processing at global scale, is its very expensive computational cost (with a large number and variety of input variables and a long processing time). A physics-aware machine learning approach can be adopted for accelerating STEMMUS-SCOPE’s running. The core idea is to approximate the original model by a surrogate machine learning model (i.e., emulator). Based on a limited number of STEMMUS-SCOPE runs, the input-output pairs (corresponding to training samples) are used to establish the emulator, which is then used to infer the model output given a yet-unseen input configuration. Currently, the Random Forest model was adopted for this purpose with the guide of physical principles (see Figure 2).
This physics-aware machine learning algorithm has been applied to produce global soil moisture products (Han et al. 2023 Sci. Data; Zhang et al. 2021 Remote Sens.), and will be applied to develop emulators for radiative transfer models (e.g., for CLAP and STEMMUS-SCOPE). The preliminary results look promising: