Towards surrogate models with hybrid spatial neural networks: a summary of results

Abstract

The goal is to develop an efficient and accurate surrogate model for Daycent, a widely used but computationally expensive ecosystem model. This problem is important due to its societal applications in sustainable agriculture. Challenges include balancing the trade-off between prediction time and solution quality (e.g., accuracy), as well as the need to capture spatial relationships both within and across sites, while also accounting for varied crop management practices that introduce irregular and non-stationary patterns, reducing predictability. Related work on surrogate models with traditional feed-forward artificial neural networks (SM-ANN) has shown that these models have limited accuracy and often fail to capture spatial dependencies. To address these limitations, we explore novel Surrogate Models with Hybrid Spatial Neural Networks (SM-Hybrid) capable of explicitly modeling spatial autocorrelation and tele-connections. Experimental results show that the proposed SM-Hybrid is more accurate than SM-ANN and is twice as fast as the Daycent model.