X-ray and neutron scattering remains one of the primary methods to obtain information on the nanostructure of lignocellulosic materials such as the wood cell wall. However, due to the complexity of the structure of the cell wall, obtaining detailed information from the scattering patterns is challenging and subject to uncertainties. Approaches based on known analytical form factors and structure factors have yielded some success. However, a more sophisticated approach relying on fewer assumptions and considering the entire 2D scattering pattern would allow to obtain more detailed information. In this work, we develop a numerical approach to generate real-space structures based on a set of structural input parameters and calculate the scattering patterns from the generated structures. We then use a machine learning approach to invert the problem to obtain the set of parameters best describing a scattering pattern. The structural parameters consist of quantities like volume fraction of crystalline microfibrils, their thickness distribution, microfibril angle, waviness descriptors, agglomeration tendency and so on. We work both on numerically simple 2D models for interpreting equatorial 1D scattering curves as well as more heavy 3D models for interpreting 2D scattering patterns. The 3D models are built with torsion-free beam elements describing fibrils undergoing fully damped Langevin dynamics to settle into a equilibrated non-interpenetrating configuration in a representative volume element. The fibril stiffness relative to the intensity of the Brownian motion determines the waviness of the fibrils and attractive interactions are used to introduce agglomeration tendencies. We work together with Paavo Penttilä's group in Aalto University and University of Jyväskylä and collaborators like the ForMAX beamline at the MAX IV synchrotron facility to design a tool to for wood researchers to infer detailed information on the nanostructure of cell walls. The goal is to release an open tool to be used by the community.