Eddy covariance (EC) is a passive, non-invasive method for measuring ecosystem-atmosphere trace gas exchange. It has become increasingly popular in recent years as hardware and software have become more accessible. Eddy covariance cannot measure fluxes continuously because the assumptions underpinning the method are not valid under all meteorologic conditions, but data from EC sites are widely used to monitor ecosystem scale energy, water, and carbon exchange. Trace gas fluxes tend to exhibit spatially and temporally variable, non-linear dependence upon numerous drivers. Multi-year EC data sets have hundreds of thousands of data points and flux time series contain both noise and data gaps. These factors make EC data poorly suited for analysis with traditional statistical methods. Here we present a guidance for leveraging the flexibility and functionality of Neural network (NN) models for working with EC data. Neural networks are a flexible machine learning method and an ideal tool for working with large multivariate datasets with complex non-linear dependencies. They offer control over the structure a model and inspection of model derivatives provides a method for ensuring that relationships mapped by a NN are physically plausible. We demonstrate methods for inferential modelling with NN and EC data, provide examples demonstrating how model derivatives can be used to detect and visualize the functional relationships, and offer comparisons to other common ML methods.