How does the machine learning model work?
Machine learning models are powerful for leveraging data, but typically require larger data than traditional modeling methods to yield optimal results. GeoFOR currently contains approximately 2,600 cases. Because data collection is ongoing, the model is updated frequently as new data is entered.
The size and comprehensive nature of the geoFOR data allow for the application of machine learning methods for estimating PMI. After users submit the details of their case, the app automates the collection of weather data and delivers a PMI prediction using a statistically robust XGBoost regression model. The cross-validating machine learning PMI predictive model results in a R² value of 0.8 and users receive a predicted PMI with an 80% confidence interval.
Model predictions vs. true PMI values
The machine learning XGBoost model is an ensemble of decision trees, similar to a flow chart. Using decomposition characteristics from cases with known PMI, the model is able to predict a PMI for additional cases.
Below is a simplified version of the model to demonstrate the intuitive relationship between decomposition characteristics and PMI. It does not incorporate variables such as weather and deposition type, thus it is not as accurate as the current PMI estimation tool provided by geoFOR.