Modeling hydrogen mobility in titanium hydrides on the basis of a single diffusivity is impossible due to the different connectivity, activation energy, and heat capacity of the face-centered cubic (FCC), body-centered cubic (BCC), and hexagonal close-packed (HCP) titanium–hydrogen compounds. This paper presents a new lattice-conditioned Arrhenius residual learning algorithm (LC-ARL) to convert predictions of the diffusion process based on lattice moment tensor potentials into the temperature dependence of the physically meaningful diffusivity value. The main problem to be solved here concerns whether a residual mapping could provide the physically meaningful ordering between Ti-H phases and which Ti-H transport phase needs further physical validation by characterizing its effective diffusivity. The developed LC-ARL algorithm includes two-dimensional evaluation of residual spaces for MTP-AL and MTP-DIRECT moment tensor potentials, determination of the phase-resolved reliability weighting factor, calculation of Arrhenius prefactor parameters using reference state diffusivities, and construction of Arrhenius diffusivity curves based on phase-resolved geometric average. The selected phase-resolved validation examples are FCC Ti648H1296, BCC Ti648H648, and dilute HCP Ti648H36. LC-ARL leads to the much lower mean absolute activation energy error (0.0084 eV) relative to MTP-AL (0.0567 eV) and MTP-DIRECT (0.0167 eV) algorithms without increasing the mean absolute logarithmic diffusivity error (identical 0.074 decade). Moreover, the temperature sweep preserves the Ti-H transport phase sequence BCC \(>\) HCP \(>>\) FCC within the 300-1000 K range. Thus, the proposed research question was positively answered concerning this particular data set – a combination of magnitude and slope errors resulted in physically meaningful Arrhenius diffusivity function. Dilute HCP titanium hydride was identified as the most appealing validation example because of non-coordinated optimization of magnitude and activation energy.
