Constrained hybrid modelling to predict microbial dynamics and organic matter turnover in soil systems

Soil microorganisms control organic matter cycling and, to a large extent, determine how soil systems can cope with and mitigate climate change and environmental threats. Integrating them explicitly into process-based soil models is critical for predicting how soil carbon (C) flows and stocks change in ecosystems with time. Models are critical tools for integrating datasets with theory. However, integrating information from modern omics-based datasets is a challenge due to the nonlinear relationship between genomes and the actual function microbes express in their local environment. Functional traits can be defined and inferred from these genomic datasets to better leverage their information and better understand the complexity of the soil microbiome. Integrating trait information with process-based microbially explicit models provides an opportunity leverage genomic data for an improved soil carbon prediction.

We present a hybrid modeling framework that uses a data-driven neural network approach to derive microbial parameters of process-based models from metagenome inferred functional traits, leveraging information from metagenomic and DNA sequencing datasets. We combine a neural network (multi-layer perceptron) with a process-based soil model to set up a hybrid model. The neural network uses genomic trait data as the input and predicts biokinetic parameters of the process-based model. We trained the hybrid model with synthetic genomic trait datasets of varying complexity and time series of state variables of the process-based model (e.g. carbon dioxide production) to demonstrate the approach. Using trait inference from genomes, the model can learn several biokinetic parameters such as growth rates, dormancy rates, affinities to organic matter, growth yields or decay rates. The training uses a complex constraint-based loss function, informing the model from ecological theory and literature data, ensuring the realistic behavior of every non observed state variable during training such as active and dormant microbial pools. Compared to a ‘naïve’ hybrid model, the use of a more complex loss function reduces model equifinality and ensure realistic behavior of the non-observed state variables. Naïve loss function cannot efficiently learn the behavior of non-observed state variables and fail to predict realistic microbial dynamics. We present i) the concept of the hybrid soil modelling framework, ii) the constraint-based loss function approach, iii) the performance of constrained versus naïve hybrid models after training with different synthetic datasets.