We present an extension to the Automated Potential Development (APD) workflow, an active-learning (AL) framework for training machine-learned interatomic potentials (MLIAPs). A mode for using pre-existing ab initio molecular dynamics (AIMD) trajectories is added, accelerating the initial data-generation stage and substantially reducing computational cost. The active-learning cycle is also accelerated through an improved defect-sampling procedure that enhances active-learning coverage. In addition, the workflow has been extended to provide compatibility with Quantum ESPRESSO, broadening the range of historical calculations that can be reused. These extensions target workflow efficiency rather than changes in the fitting objective or target accuracy. The resulting potentials are intended for applications within the finite-temperature regime sampled by the reused AIMD data. Applied to two oxide-ion conductors, Bi 0.852 V 0.148 O 1.648 and Bi 0.852 P 0.148 O 1.648 , it enables multi-nanosecond simulations across a wide temperature range while requiring an order of magnitude fewer new DFT calculations. In Bi 0.852 V 0.148 O 1.648 , the method reproduces high-temperature DFT results and captures a low-temperature pre-diffusive regime consistent with quasielastic neutron scattering and nuclear magnetic resonance observations. Leveraging existing DFT data within this AL framework thus offers an efficient route to developing accurate MLIAPs for complex ionic conductors.
Duff et al. (Sun,) studied this question.