Skip to main content
Photon Sciences

Machine-learning Determines Interatomic Potentials from X-ray Diffraction Data

A new machine learning technique that uses data from high-energy X-ray diffraction experiments will significantly reduce model development time and human effort.

Understanding the structure and properties of refractory oxides is critical for high-temperature applications. Their high melting temperatures (> 1500 °C) make refractories suitable for applications in harsh environments, in addition to their insulating properties and ability to prevent oxidation. It is therefore important to identify phase transformations and structural rearrangements close to the melting point. Using a machine learning methodology based on high-energy X-ray data obtained from laser heating, a multiphase potential is generated for a canonical example of the archetypal refractory oxide, HfO2, by drawing a minimum number of training configurations from room temperature to the liquid state at ∼2900 °C. The method significantly reduces model development time and human effort.

Our high-energy photon diffraction and spallation neutron Pair Distribution Function (PDF) measurements provide the starting point for a deep insight into the structure of disordered materials through machine learning. In this new ML scheme, the experimental PDF data drives an active learning algorithm that tests ab initio molecular dynamics (AIMD) simulations using a Gaussian Approximation Potential (GAP) approach. The method uses an automated closed loop via an active-learner, which is initialized by X-ray and neutron diffraction measurements, and sequentially improves an unsupervised machine-learning model until the experimentally predetermined phase space is covered. The resulting classical GAP MD simulations reproduce all the experimental phases with near ab initio precision and yield quench rates of 1K/ps not accessible via AIMD.