Constructing and evaluating machine-learned interatomic potentials for Li-based disordered rocksalts

Lithium-based disordered rocksalts (LDRs), which are an important class of cathodes for advanced Li-ion batteries, represent a complex chemical and configurational space for conventional density functional theory (DFT)-based high-throughput screening approaches. Notably, atom-centered machine-learned interatomic potentials (MLIPs) are a promising pathway to accurately model the potential energy surface of highly-disordered systems, such as LDRs, where the performance of such MLIPs have not been rigorously explored yet. Here, we represent a comprehensive evaluation of the accuracy, transferability, and ease of training of five MLIPs in modelling LDRs, including artificial neural network potential developed by the atomic energy network (AENET), Gaussian approximation potential (GAP), spectral neighbor analysis potential (SNAP) and its quadratic extension (qSNAP), and moment tensor potential (MTP). Specifically, we generate a DFT-calculated dataset of 10842 disordered LiTMO$_2$ and TMO$_2$ configurations, where TM = Sc, Ti, V, Cr, Mn, Fe, Co, Ni, and/or Cu. Importantly, we find AENET to be the best in terms of accuracy and transferability for energy predictions, while MTP is the best for atomic forces. While AENET is the fastest to train at low number of epochs, the training time increases significantly as epochs increase, with a corresponding reduction in training errors. Note that AENET and GAP tend to overfit in small datasets, with the extent of overfitting reducing with larger datasets. Finally, we observe AENET to provide reasonable predictions of average Li-intercalation voltages in layered, single-TM LiTMO$_2$ frameworks, compared to DFT ($\sim$10% error on average). Our study should pave the way both for discovering novel LDR electrodes and for modelling other configurationally complex systems, such as high-entropy ceramics and alloys.