The models that performed the best, and so we chose to highlight, were multilayer perceptrons, kernel support vector regressor and both gradient boosting methods. While some of these models were highlighted in existing literature, some were the results of our own exploration. These top-performing models had an R2 score of at least 0.81. While they are standalone models that can be individually used, we developed an interface that brings them together in an easy-to-use methane emission calculator.
This methane emission calculator interface allows users to input and customize cattle-specific information to generate daily methane emission predictions using the selected model. Users can also opt to run all the models simultaneously for comparison. In particular, the interface takes feed components and the relative abundance of microbiome organisms as input parameters.
Using this methane emission calculator, we can estimate our solution’s effect on cattle. If we know how our solution affects the rumen’s microbiome, we can see how methane emissions would be affected. Outside the scope of BovEco, this calculator can simply be used by cattle farmers who may seek to optimize feed to minimize methane emissions.
For data visualization, please feel free to look at the outside_cattle.ipynb in the scripts folder. To explore our interface, clone the repository and follow these steps:
-
Make sure to install git lfs by running
git lfs install. If any issues occur, refer to the website here: https://git-lfs.com/ -
Make sure install a MATLAB version compatible with python version
pip install matlabengine -
Make sure you are in the right directory by running
cd models/outside-cattle/scripts -
Install the required packages by running
pip install -r requirements.txt - Run the interface by running
streamlit run main.py
This is the interface at localhost:8501
Additionally, as a means of possibly connecting the feed intake of the cattle to the methane emission, an interface has been developed for the previously mentioned ODE system that deduce concentrations of polymer (feed) components, soluble components, microbial groups and amount of components in the gas phase before and after introduction of PeIR into the system.
Similarly to the above, follow the steps 1, 2, 5, replace step 3 and 4
with: cd models/inside-cattle/scripts
The interface will look like the figure below:
We hope that further development of such an interface can provide an avenue into providing an estimation of the effectiveness of methane mitigation strategies, more specifically through feed formulation in our case. Through an iterative software process pipeline, consisting of incorporating the inputs from a mechanistic model of the rumen and translating that to a predicted methane emission through a trained machine learning model, this may provide direct insights to farmers as to how to feed their livestock in an optimally eco friendly way, and reduce overfeeding in the process.