Abstract
Ice templating creates anisotropic porous materials whose pore architecture depends on freezing-front dynamics. In cryostage-based processing, the manipulated input is reference temperature rather than freezing-front position. Fixed setpoints and constant cooling rates do not directly prescribe a desired front trajectory, while programmed profiles must account for the realised plate response and transient phase-change dynamics. In the present thermocouple-instrumented configuration, front position is not continuously available for feedback, motivating an open-loop inverse-design approach.We develop a physics-based framework using Gaussian-process Bayesian optimisation to determine executable reference-temperature trajectories for approximately constant-velocity centreline front progression. Each candidate is evaluated using a calibrated reference-to-plate model coupled to an axisymmetric finite-element phase-change simulation, while Bayesian optimisation guides the search over bounded, monotone trajectories against a linear front-position target.For purified water, the phase-change model reproduced post-fill thermocouple measurements at three heights with a mean absolute error of 0.72 °C and a root-mean-square error of 0.91 °C. Full-resolution confirmation for target mean centreline freezing-front velocities of 6–10 μm s-1 yielded mean-speed errors from +1.41% to -2.38% and front-position root-mean-square errors of 0.039–0.176 mm over the 2.5–11.5 mm tracking interval. Experimental validation at 6, 7, and 8 μm s-1 reproduced the predicted lower-to-upper freezing sequence. Measured 0 °C crossing times at 3.0 and 6.2 mm differed from predictions by less than 0.8 min, whereas freezing occurred 1.3–2.6 min earlier at 11.0 mm. These results demonstrate that physics-based Bayesian inverse design can translate a desired freezing-front trajectory into executable cryostage temperature trajectories.