From the supplementary of [1], with some care should be applicable to the NEB / minimization as well.
Maybe at the potential level itself, as I understand this doesn't involve multiple frame averaging, just a random rotation on each structure as it comes (with the inverse rotation for the forces of course) and this apparently helps a lot, as heard at Michele's talk at the CECAM workshop on atomistic machine learning software.
[1]: M. F. Langer, S. N. Pozdnyakov, and M. Ceriotti, “Probing the effects of broken symmetries in machine learning,” Mach. Learn.: Sci. Technol., vol. 5, no. 4, p. 4LT01, Dec. 2024, doi: 10.1088/2632-2153/ad86a0.
From the supplementary of [1], with some care should be applicable to the NEB / minimization as well.
Maybe at the potential level itself, as I understand this doesn't involve multiple frame averaging, just a random rotation on each structure as it comes (with the inverse rotation for the forces of course) and this apparently helps a lot, as heard at Michele's talk at the CECAM workshop on atomistic machine learning software.
[1]: M. F. Langer, S. N. Pozdnyakov, and M. Ceriotti, “Probing the effects of broken symmetries in machine learning,” Mach. Learn.: Sci. Technol., vol. 5, no. 4, p. 4LT01, Dec. 2024, doi: 10.1088/2632-2153/ad86a0.