Abstract: Extracting accurate models for complex phases in material from experimental data is extremely challenging. Even though scattering data are rich in microscopic information, the conventional approaches struggle to deal with the scale and complexity.  

A model-based machine-learning (ML) technique was used to address the ill-pose nature of the inverse scattering problem. Neural networks were trained on Monte-Carlo simulations for numerous microscopic models sampled over the parameter space using an iterative algorithm. The trained networks can learn to detect and extract the model’s key features by filtering out irrelevant information like noise and experimental artifacts by projecting experimental data on a finite-dimensional space determined by the NN architecture. Further, the compressed representation of data can be used to classify regions with different properties, generate phase diagrams and build fast Surrogates to bypass expensive calculations. This approach is shown to provide a better understanding of the formation of a glass on cooling the spin liquid Dy2Ti2O7, quantum spin liquids on honeycomb lattices, and understanding the interactions and phases in RuCl3. Examples of them as well as powder scattering data are shown.