Surrogate representation of sink strengths and the long-term role of crystalline interfaces in the development of irradiation-induced bubbles

The present article addresses an early-stage attempt on replacing the analyticity-based sink strength terms in rate equations by surrogate models of machine learning representation. Here we emphasise, in the context of multiscale modelling, a combinative use of machine learning with scale analysis, through which a set of fine-resolution problems of partial differential equations describing the (quasi-steady) short-range individual sink behaviour can be asymptotically sorted out from the mean-field kinetics. Hence the training of machine learning is restrictively oriented, that is, to express the local and already identified, but analytically unavailable nonlinear functional relationships between the sink strengths and other local continuum field quantities. With the trained models, one is enabled to quantitatively investigate the biased effect shown by a void/bubble being a point defect sink, and the results are compared with existing ones over well-studied scenarios. Moreover, the faster diffusive mechanisms on crystalline interfaces are distinguishingly modelled by locally planar rate equations, and their linkages with rate equations for bulk diffusion are formulated through derivative jumps of point defect concentrations across the interfaces. Thus the distinctive role of crystalline interfaces as partial sinks and quick diffusive channels can be investigated. Methodologicalwise, the present treatment is also applicable for studying more complicated situation of long-term sink behaviour observed in irradiated materials.