• Accepted Paper

Data-driven test of local recoverability for materials design

Zachary T. Varley, Megna Shah, Veera Sundararaghavan, and Jeff P. Simmons

Phys. Rev. E - Accepted 3 November, 2026

DOI: https://doi.org/10.1103/j6cn-ybsv

Abstract

Materials design often requires identifying processing conditions that produce a target microstructure. When microstructures are represented by continuous descriptors, inverse design becomes a navigation problem: can a local region of descriptor space be mapped back to a coherent neighborhood of processing conditions? We refer to this property as local recoverability. It can fail when distinct processes produce overlapping descriptors, when descriptor variation contracts relative to control variation, or when descriptors saturate. We present a data-driven, pointwise diagnostic: within outcome-space neighborhoods, a ridge-stabilized local inverse is fit and scored by leave-one-out explained variation. The resulting score identifies where inverse recovery is reliable and where it breaks down, without requiring gradients, adjoints, or a mechanistic model. We demonstrate the method on synthetic mappings constructed to exhibit each of these failure modes by design, on Cahn–Hilliard phase-field simulations under both deterministic and ensemble acquisition, and on three-state Potts lattice simulations with ensemble-averaged correlation descriptors. In each case the score localizes the failure regions predicted by the construction (e.g., the folded region of the Cahn–Hilliard control plane and the majority-state-saturated region of the Potts control plane) and leaves the rest of the parameter domain flagged as recoverable.

Export citation

Export citation

Choose format for download:

Download Citation

If the author has provided any supplemental materials with this article they will be available upon publication of the version of record.

Sign In to Your Journals Account

Filter

Filter

Article Lookup

Enter a citation