Performance-Portable Implementation of SISSO++ and its Application in Materials’ Mobility Prediction

Sure-independence screening and sparsifying operator (SISSO) is a powerful, artificial intelligence tool for identifying symbolic descriptors and predictive models [1]. It has been successfully used to discover new optimal materials in a number of applications, e.g. thermal conductivity [2]. Recently, we have developed SISSO++, a new implementation of SISSO that uses both OpenMP and MPI and achieves scalable parallel performance on CPU clusters [3]. Here, we present an updated implementation that uses the C++ performance-portability framework Kokkos [4] to offload the performance-critical regions of the algorithm to accelerators, such as Nvidia or AMD GPUs. With the SISSO++ code, we trained models for the experimentally measured charge carrier mobility of bulk materials. Primary features used for training are derived from first principles, including structural, electronic structure, vibrations, and electron-vibrational coupling properties of the materials. We then discuss how these primary features influence carrier mobilities, shedding light on the underlying mechanisms governing materials' electronic transport properties.

 

[1] R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, et al. Phys. Rev. Mater. 2, 083802 (2018).

[2] T. Purcell, M Scheffler, L. Ghiringhelli, C. Carbogno. npj Comput. Mater. 9, 112 (2023).

[3] T. Purcell, M. Scheffler, C. Carbogno, L. Ghiringhelli. J. Open Source Softw. 7, 3960 (2022).

[4] C. Trott, D. Lebrun-Grandié, D. Arndt. J. Ciesko, et al. IEEE Transactions on Parallel and Distributed Systems 33, 805 (2022).