A deep learning approach for quantum dots sizing from wide-angle X-ray scattering data

Disclosing the full potential of functional nanomaterials requires the optimization of synthethic protocols and an effective size screening tool, aiming at efficiently triggering their size-dependent properties. Here we demonstrate the successful combination of a wide-angle X-ray total scattering approach with a deep learning classifier for directly sizing quantum dots in both colloidal and dry states. This work offers a compelling alternative to the lengthy process of deriving sizing curves from transmission electron microscopy coupled with spectroscopic measurements, especially in the ultra-small size regime, where traditional empirical functions exhibit larger discrepancies. The core of our algorithm is an all-convolutional neural network trained on Debye scattering equation X-ray simulations, incorporating atomistic models to capture structural and morphological features, and augmented with physics-informed perturbations to account for different predictable experimental conditions. The model performances are evaluated using both wide-angle X-ray total scattering simulations and experimental datasets collected on lead sulfide quantum dots, resulting in size classification accuracies surpassing 97%. With the developed deep learning size classifier, we overcome the need for calibration curves for quantum dots sizing and thanks to the unified modeling approach at the basis of the total scattering method implemented, we include simultaneously structural and microstructural aspects in the classification process, This algorithm can be complemented by incorporating input information from other experimental observations (e.g. small angle X-ray scattering data) and can be easily extended to other classes of nanocrystals, providing the nanoscience community with a powerful and broad tool to accelerate the development of novel functional (nano)materials.