InvDesMobility: a reliability-gated first-principles feedback framework for closed-loop materials discovery

A research team has detailed InvDesMobility, a closed-loop materials discovery framework that gates expensive first-principles calculations behind automated reliability checks before feeding results back into generative and acquisition models [1]. The framework, described in a preprint posted to arXiv on June 15, 2026, targets composite properties such as carrier mobility, where a final scalar value can obscure intermediate quantities, fit quality, and convergence history [1]. InvDesMobility integrates multi-agent automated density functional theory (DFT), evidence stratification, generative structure proposal, acquisition ranking, and auditable release [1]. The authors argue that the value of inverse design in closed-loop discovery depends on whether expensive first-principles results are independently validated, provenance-recorded, and admitted as feedback only when evidence is sufficient [1]. The workflow began with 516 candidates drawn from the 2DMatPedia database [1]. After quality-control filtering, 280 materials passed, yielding 573 retained carrier-direction seed channels following channel-level reliability gating [1]. Those records were split into two feedback objects: relaxed structures updated the generative model, while retained mobility channels trained the acquisition model and set validation priority [1]. Over multiple iterations, InvDesMobility screened 2.4 x 10^6 structures and submitted 102 candidates for DFT validation [1]. The process retained 86 reliability-gated generated channels across 41 distinct formulas [1]. The authors emphasize that the primary contribution is not a fixed list of high-mobility materials but a transferable feedback contract that makes closed-loop inverse design both useful and auditable when learning from expensive calculated properties [1]. All source data, retained feedback records, and workflows have been released on GitHub, accompanied by an evidence website [1]. The preprint appears on arXiv, an open-access repository that, as of November 2024, receives roughly 24,000 submissions per month and hosts more than two million articles across physics, mathematics, computer science, and related fields [6]. arXiv itself notes that e-prints are moderated but not peer-reviewed [6]. The platform also supports arXivLabs, a framework for community-contributed tools such as bibliographic explorers and code linkers, though new project proposals are temporarily paused while the development team focuses on cloud migration [3][4][5].