Autonomous crystal structure search by crystal morphing

Recently, the autonomous search of materials was studied to alleviate the costs of material developments. For the crystalline solid domain, the crystal invariances are essential for an efficient autonomous search. If the invariances are not promised in the search space, myriad redundant points that represent the same crystal structure exist, which obstructs a construction of an efficient crystal-search space.

In this talk, we propose the crystal morphing to construct the crystal-invariant search space based on the reciprocal space SOAP distance, and its application to an x-ray crystallography. We use the analytical formula of derivatives of the SOAP distance together with the gradient method to implement the crystal morphing. Contrary to the other generative models based on the deep-neural network schemes, our method mathematically implements the invariances without any learning steps, allowing the creation of an efficient and reliable crystal-search space from inputs of known structures.

We will show how the crystal morphing works and the result of the demonstration of x-ray crystallography. We will discuss the potential usefulness of the crystal morphing that automatically found a better initial guess for the Rietveld analysis by inputting known structures.