In [1912.11055] we proposed to replace the exact squared-amplitudes used in monte carlo (MC) event generators with approximate, albeit very precise, ones in the form of pre-trained machine learning (ML) regressors. The idea is to speed up the evaluation of the numerically expensive functions that arise in loop computations. This approach also alleviates the need for quadruple and higher precision arithmetic during event generation. In this talk I will start by discussing a proof of principle that demonstrates the efficacy of this proposal. In the rest of the talk, I will discuss our progress towards the ultimate goal of approximating building blocks of NNLO squared-amplitudes where the gain in evaluation speed can be even more dramatic.
