Ultra-low field MR systems rely on bulky and expensive shielding enclosures that attenuate the large ambient electromagnetic ambient noise intrinsic to this regime. This increases system costs and hinders portability. To reduce this shielding constraint we propose using a software gradiometer based on convolutional neural networks. The presented approach employs three peripheral coils to estimate the ambient noise interfering with the NMR acquisition. Preliminary results suggest that this method can provide a significant noise reduction. Such a system could considerably reduce shielding requirements promoting the system’s portability and reducing its costs.