lambda_tau
This keyword sets the weight of spin-torque RMSE in the loss:
lambda_tau <weight>
The weight must be non-negative and defaults to 0. For each labeled active atom, the predicted and target quantities are converted to
The resulting torque RMSE emphasizes magnetic-force components perpendicular
to the current spin. Only frames containing mforce:R:3 and atom types
enabled by spin_type contribute.
The Cartesian loss lambda_m constrains all magnetic-force
components, whereas torque is insensitive to the parallel component.
Set lambda_m 0 and a positive lambda_tau to supervise spin rotation
without constraining the longitudinal magnetic force. Torque RMSE includes only
labeled active spins with \(|\boldsymbol{s}_i|^2 > 10^{-20}\,\mu_\mathrm{B}^2\)
and is normalized by two degrees of freedom per included spin. The squared
magnitude is evaluated in double precision for both the error and its count;
spins at or below the threshold are excluded from both. Squared torque errors
weight squared transverse magnetic-force errors by \(|\boldsymbol{s}_i|^2\).