Hybrid AI Models for Materials Discovery with Versioned Laboratory Datasets: A Reproducibility-Centred Evaluation

Abstract

Materials-screening datasets often mix experimental batches, computed descriptors and inconsistent processing metadata. The difficulty is particularly visible when a system is transferred to a new setting. To address this, the study will examine whether hybrid prediction models remain reliable when material families and laboratory batches are held out. The research design centres on a reproducibility-focused benchmark built from versioned materials datasets and documented processing histories. The choice of evaluation units and exclusion criteria is informed by the broader methodological record [3–7]. The analysis will contrast physical-descriptor baselines, purely learned models and hybrid models combining domain constraints with machine learning. The reproducibility extension records dataset versions, model settings, exclusion rules and provenance before any held-out evaluation. Related enterprise-storage studies inform the assessment of operational reliability [1] and the deployment assumptions adopted for system-level comparison [2]. Assessment will focus on out-of-family prediction error, calibration and stability of candidate rankings under data revisions. The supporting evidence base motivates the chosen comparator conditions and robustness criteria [8–12]. Critical sensitivity checks address recipe duplicates, provenance gaps, leakage across related materials and sensitivity to batch-specific artefacts. Further systems and modelling literature supplies context for the proposed sensitivity analyses [13–17]. An independent replication exercise would repeat the main analysis from the documented artefacts and quantify disagreements. The expected output is a reusable evaluation framework with transparent uncertainty and reporting rules, informed by wider reproducibility literature [18–23].