Opportunistic emulation of computationally expensive simulations via Deep Learning

Conrad Sanderson, Dan Pagendam, Brendan Power, Frederick Bennett, Ross Darnell

Abstract

With the underlying aim of increasing efficiency of computational modelling pertinent for managing and protecting the Great Barrier Reef, we perform a preliminary investigation on the use of deep neural networks for opportunistic model emulation of APSIM models, by repurposing an existing large dataset containing the outputs of APSIM model runs. The dataset has not been specifically tailored for the model emulation task. We employ two neural network architectures for the emulation task: densely connected feed-forward neural network (FFNN), and gated recurrent unit feeding into FFNN (GRU-FFNN), a type of a recurrent neural network. Various configurations of the architectures are trialled. A minimum correlation statistic is employed to identify clusters of APSIM scenarios that can be aggregated to form training sets for model emulation.

Introduction

Protecting the Great Barrier Reef requires a deep understanding of how various farming practices impact water quality. Scientists traditionally rely on the Agricultural Production Systems sIMulator (APSIM) to model these complex interactions. However, APSIM simulations are computationally expensive; running millions of "what-if" scenarios to inform policy can take months of processing time.

To address this bottleneck, researchers are developing emulators—surrogate models that act as mathematical shortcuts. These emulators mimic the complex simulator's behavior but run nearly instantaneously. This study explores "opportunistic emulation": the ability to build these models by repurposing existing datasets from past projects, rather than generating expensive, perfectly tailored new data.

Methodology

This research compared two primary Deep Learning architectures to determine their efficacy in handling the time-series nature of agricultural data:

  1. Feed-Forward Neural Networks (FFNN): A standard architecture that processes each day's data independently, lacking temporal context.
  2. GRU-FFNN: A hybrid model utilizing Gated Recurrent Units (GRUs). As a type of Recurrent Neural Network (RNN), GRUs are designed to "remember" information from previous time steps, which is critical for modeling soil and water states that depend on antecedent weather and management conditions.
The GRU-FFNN architecture incorporates recurrent connections to maintain state across time steps (t-1, t, t+1), allowing the model to account for historical environmental data.

Key Innovations

  • Feature Engineering: Beyond raw weather data, the team created 26 derived features, such as "smoothed" rainfall and fertilizer application metrics, to provide the models with long-range temporal context.
  • Clustering via Correlation: To ensure training consistency, a "minimum correlation" statistic was used to group similar simulation scenarios.
  • Funnel Architecture: A "conical" design was implemented for the feed-forward layers, halving the number of nodes in each subsequent layer to prevent overfitting and memorization.

Results and Discussion

The study utilized the Paddock to Reef (P2R) dataset, comprising decades of APSIM simulation runs across various soil types and management practices. The models targeted four key outputs: runoff, soil_loss (erosion), DINrunoff (dissolved inorganic nitrogen), and Nleached (nitrogen leaching).

Comparison of emulation performance between FFNN (top) and GRU-FFNN (bottom) across four target variables. The GRU-FFNN shows significantly tighter correlation for runoff and DINrunoff.

The Importance of Memory

The GRU-FFNN significantly outperformed the standard FFNN, particularly for runoff and dissolved nitrogen (DINrunoff). This confirms that accurate agricultural emulation requires a "memory" of previous soil moisture and nutrient levels. However, more data did not always lead to better results; lowering similarity thresholds to include more training data sometimes decreased accuracy, suggesting that the added data introduced conflicting dynamics or "noise."

The "Opportunistic" Limitation

A critical finding was that large datasets are not necessarily "rich" datasets. Because the P2R data was collected for other purposes, it contained many "zeros" (dry periods) and lacked the diversity needed to teach the neural networks about rare, high-impact events. Consequently, all models struggled to capture the high-variability "spikes" in soil_loss and Nleached, which are often the most critical events for environmental management.

Conclusion

While opportunistic data is sufficient for emulating relatively linear processes like water runoff, it reaches its limit with complex dynamics like soil erosion. Future work should shift toward Design of Computer Experiments (DoCE). By generating richer datasets using synthetic, 1,000-year weather files, researchers can provide Deep Learning models with the dense examples of weather extremes necessary to learn the full range of environmental behavior. Once perfected, these emulators can reduce months of simulation work into hours, drastically accelerating reef protection efforts.

Summary updated

Abstract

With the underlying aim of increasing efficiency of computational modelling pertinent for managing and protecting the Great Barrier Reef, we perform a preliminary investigation on the use of deep neural networks for opportunistic model emulation of APSIM models, by repurposing an existing large dataset containing the outputs of APSIM model runs. The dataset has not been specifically tailored for the model emulation task. We employ two neural network architectures for the emulation task: densely connected feed-forward neural network (FFNN), and gated recurrent unit feeding into FFNN (GRU-FFNN), a type of a recurrent neural network. Various configurations of the architectures are trialled. A minimum correlation statistic is employed to identify clusters of APSIM scenarios that can be aggregated to form training sets for model emulation.
BibTeX
@inproceedings{sanderson2021opportunistic,
  author = {Conrad Sanderson and Dan Pagendam and Brendan Power and Frederick Bennett and Ross Darnell},
  title = {{Opportunistic emulation of computationally expensive simulations via Deep Learning}},
  booktitle = {MODSIM2021,},
  year = {2021},
  doi = {10.36334/modsim.2021.l1.sanderson},
}
Figure