Reply on RC2

Chaopeng ShenOpenalex Unknown 2023 · DOI · OpenAlex

Abstract

Photosynthesis plays an important role in carbon, nitrogen, and water cycles. Ecosystem models for photosynthesis are characterized by many parameters that are obtained from limited in-situ measurements and applied to the same plant types. Previous site-by-site calibration approaches could not leverage big data and faced issues like overfitting or parameter non-uniqueness. Here we developed a programmatically differentiable version of the photosynthesis process representation within the Functionally Assembled Terrestrial Ecosystem Simulator (FATES) model. This model is coupled to neural networks that learn parameterization from observations of photosynthesis rates. We first demonstrated that the framework was able to recover multiple assumed parameter values concurrently using synthetic training data. Then, using a real-world dataset consisting of many different plant functional types, we learned parameters that performed substantially better and dramatically reduced biases compared to literature values. Further, the framework allowed us to gain insights at a large scale. Our results showed that the carboxylation rate at 25 °C ($V_{c,max25}$), was more impactful than a factor representing water limitation, although tuning both was helpful in addressing biases with the default values. This framework could potentially enable a substantial improvement in our capability to learn parameters and reduce biases for ecosystem modeling at large scales.

Introduction

Predicting how vegetation interacts with the atmosphere is fundamental to climate science. However, mechanistic ecosystem models are notoriously difficult to calibrate. These models rely on numerous parameters, such as the maximum rate of carboxylation ($V_{c,max25}$), which are typically derived from sparse, local measurements and then extrapolated across broad Plant Functional Types (PFTs).

Traditional site-by-site calibration methods are computationally intensive and often lead to "equifinality"—a state where multiple parameter combinations produce similar outputs, making it difficult to identify physically accurate values. As ecological "big data" becomes more available, there is an urgent need for scalable frameworks that can leverage global datasets to improve model accuracy.

Methodology

This study introduces a hybrid approach termed Differentiable Parameter Learning (dPL). The researchers re-engineered the photosynthesis module of the Functionally Assembled Terrestrial Ecosystem Simulator (FATES) to be "programmatically differentiable." This allows the model's physical equations to be integrated directly into the backpropagation algorithms used to train neural networks.

The Hybrid Architecture

The framework utilizes a Multi-Layer Perceptron (MLP) to predict physical parameters based on inputs like plant type and soil properties. These predicted parameters are then passed into the differentiable FATES equations to calculate photosynthesis rates. By making the entire pipeline differentiable, the system can be optimized using standard gradient descent.

Implementation and Data

The system was implemented using Julia and PyTorch, utilizing parallel Newton iterations on GPUs to achieve significant speedups. Training on thousands of data points can be completed in under 10 minutes. The model was validated using:

  • Synthetic Data: To ensure the framework could accurately recover known parameter values.
  • Lin15 Dataset: A global leaf gas exchange database with thousands of measurements from 43 sites.
  • ERA5 Reanalysis & Soil Data: Used for meteorological forcing and estimating soil water stress.

Key Insights

A critical finding was the framework's ability to achieve simultaneous identifiability. It successfully distinguished between biological factors (like $V_{c,max25}$) and environmental stressors (like soil water limitation) because these factors influence the photosynthesis curve in distinct ways.

The results highlighted the dominance of biology; tuning $V_{c,max25}$ proved to be a more effective lever for reducing simulation bias than adjusting soil water parameters. Furthermore, by training on all sites simultaneously, the model benefited from data synergy, learning generalizable rules about plant behavior across diverse climates rather than overfitting to specific locations.

Results

The dPL framework demonstrated substantial improvements over traditional modeling approaches:

  • Bias Reduction: In real-world applications, bias in photosynthesis simulations dropped from 1.476 $\mu mol/m^2/s$ (literature values) to 0.347 $\mu mol/m^2/s$.
  • Increased Accuracy: The Pearson correlation coefficient (COR) improved from 0.565 to 0.748.
  • Correcting C4 Grasses: The model identified that standard literature values often overestimate carboxylation rates for C4 grasses. The learned, lower values effectively resolved a long-standing overestimation bias in the FATES model for these species.

Limitations and Future Directions

While promising, the framework has limitations. Performance was slightly lower on temporal holdout tests, suggesting that the model may be missing time-dependent factors like $CO_2$ fertilization or long-term adaptation. Additionally, data gaps for certain plant types, particularly C4 plants, remain a challenge.

Future work may involve:

  1. Neural-PBM Integration: Using neural networks to replace entire uncertain physical modules rather than just tuning parameters.
  2. Ecological Memory: Incorporating recurrent architectures (e.g., LSTMs) to account for the legacy effects of past environmental conditions, such as droughts.

Conclusion

The development of a differentiable version of FATES represents a significant step forward in ecosystem modeling. By marrying deep learning with physical process representation, this framework enables the efficient use of large-scale datasets to reduce model biases and improve our understanding of global carbon and water cycles.

Summary updated

Abstract

<strong class="journal-contentHeaderColor">Abstract.</strong> Photosynthesis plays an important role in carbon, nitrogen, and water cycles. Ecosystem models for photosynthesis are characterized by many parameters that are obtained from limited in-situ measurements and applied to the same plant types. Previous site-by-site calibration approaches could not leverage big data and faced issues like overfitting or parameter non-uniqueness. Here we developed a programmatically differentiable (meaning gradients of outputs to variables used in the model can be obtained efficiently and accurately) version of the photosynthesis process representation within the Functionally Assembled Terrestrial Ecosystem Simulator (FATES) model. This model is coupled to neural networks that learn parameterization from observations of photosynthesis rates. We first demonstrated that the framework was able to recover multiple assumed parameter values concurrently using synthetic training data. Then, using a real-world dataset consisting of many different plant functional types, we learned parameters that performed substantially better and dramatically reduced biases compared to literature values. Further, the framework allowed us to gain insights at a large scale. Our results showed that the carboxylation rate at 25 &deg;C (V<sub>c,max25</sub>), was more impactful than a factor representing water limitation, although tuning both was helpful in addressing biases with the default values. This framework could potentially enable a substantial improvement in our capability to learn parameters and reduce biases for ecosystem modeling at large scales.
BibTeX
@inproceedings{shen2023reply,
  author = {Chaopeng Shen},
  title = {{Reply on RC2}},
  booktitle = {Openalex Unknown},
  year = {2023},
  doi = {10.5194/bg-2022-211-ac1},
}
Figure