A computational framework termed FidelitySwitch-2D has demonstrated that combining candidate-specific risk estimation, adaptive routing, and group-level conformal triage can systematically curb negative transfer when translating electronic property predictions across differing materials repositories. Reported by researcher Amira Samy Talaat in Scientific Reports, the investigation addresses a persistent challenge in high-throughput materials discovery: electronic-structure databases frequently assign conflicting property values to identical two-dimensional materials due to divergent underlying density-functional theory workflows. When flexible statistical corrections are fitted to reconcile these repository gaps, they risk magnifying route-specific discrepancies when deployed outside their development environment. In prediction-sealed external evaluations across four hundred and ninety independent material groups from the MC2D database, the frozen routing framework successfully mitigated the negative transfer of a nonlinear residual correction, trimming mean absolute error by 0.0557 electronvolts. The protocol establishes an auditable reliability layer that decides which correction complexity to trust and flags high-risk material candidates before automated screening decisions proceed.
High-throughput electronic-structure repositories have dramatically accelerated virtual materials exploration by accumulating quantum-mechanical simulations for tens of thousands of candidate monolayer crystals. However, the numerical outputs cataloged within repositories such as the Computational 2D Materials Database, the JARVIS-DFT two-dimensional subset, and 2DMatPedia represent specific computational implementations rather than workflow-independent physical constants. These databases diverge fundamentally across multiple computational parameters, encompassing initial structural generation, atomic relaxation tolerances, exchange-correlation functional approximations, van der Waals dispersion corrections, spin-orbit coupling configurations, and diverse metadata conventions. Consequently, an extracted band gap constitutes an observation bound to a specific material-workflow pair, causing identical crystalline monolayers to register substantially disparate band gaps across separate repositories. In practical high-throughput screening campaigns, such numerical discrepancies can distort whether a two-dimensional candidate is categorized as insulating or semiconducting, whether its band gap falls inside a desired photovoltaic absorption window, and whether expensive experimental synthesis or higher-level calculations are ultimately warranted.
To establish an auditable benchmark that avoids statistical data leakage across related crystal records, the research team implemented a rigorous structural identity graph and standardized matching protocol. Candidate two-dimensional monolayers were normalized to a twenty-five angstrom out-of-plane vacuum spacing to prevent spurious interlayer interactions before undergoing geometric screening based on formula-unit area, in-plane lattice shape, unit cell angles, and supercell transformation ratios. Using Pymatgen StructureMatcher parameter profiles, candidate pairings were evaluated under strict criteria requiring length tolerances within ten percent, site tolerances within twenty percent, and angle tolerances within three degrees, alongside a standard tier permitting twenty percent length, thirty percent site, and five-degree variations. Lenient matches were entirely discarded, yielding an accepted graph of 1,416 strict and 556 standard cross-source pairings organized into 1,056 distinct material groups. By grouping all associated cross-source and duplicate structural observations into singular connected material entities, the researchers ensured that linked crystalline representations remained strictly partitioned together within five outer cross-validation folds, preventing row-level training data from artificially inflating downstream evaluation metrics.
The benchmark formulated cross-database translation through a directed paired structure, establishing four core baseline models of increasing complexity to map property values across disparate simulation workflows. The baseline hierarchy began with an Identity transformation preserving the original source value, followed by a Constant Delta formulation providing an additive route-level average adjustment. A more expressive Huber Anchor baseline fitted robust route-specific affine transformations, while the most complex candidate, termed the Residual Bridge, augmented the Huber Anchor with an ExtraTrees nonlinear residual model restricted by a clipping window between negative 0.8 and positive 0.8 electronvolts. To suppress unwarranted overfitting, the Residual Bridge was activated only when inner cross-validation demonstrated a material-group performance improvement of at least 0.0025 electronvolts over the affine baseline. Across 4,540 directed development pairs spanning the 1,056 material groups, the Residual Bridge achieved a development material-group macro mean absolute error of 0.2465 electronvolts, delivering a modest paired improvement of 0.0027 electronvolts over the Huber Anchor benchmark.
To govern when to deploy complex nonlinear corrections versus simpler transformations, the team designed the RiskArgminRouter and its deployable implementation, FidelitySwitch-2D. The system trains candidate-specific absolute error estimators directly on out-of-fold development residuals, evaluating each incoming material observation through predicted mean error, model uncertainty, and designated complexity penalties. When rival candidate models demonstrate comparable predicted performance within a calibrated simplicity margin, the router enforces a strict inductive preference favoring simpler candidates in the fixed order of Identity, Huber Anchor, and finally Residual Bridge. In Stage 7A development benchmarks, the RiskArgminRouter achieved the lowest non-oracle macro mean absolute error at 0.2393 electronvolts, outperforming a standard learned gate at 0.2397 electronvolts, the frozen FidelitySwitch-2D implementation at 0.2401 electronvolts, and a soft mixture-of-experts model at 0.2402 electronvolts. Meanwhile, a hypothetical oracle router selecting the optimal candidate retrospectively attained 0.2215 electronvolts, confirming significant structural headroom for adaptive model arbitration.
The critical utility of prediction-sealed routing emerged during frozen external audits designed to simulate real-world database transport. On an audit of 410 current C2DB material groups, the highly expressive Residual Bridge exhibited overt negative transfer: its error expanded to 0.3193 electronvolts, underperforming both the simple Identity baseline at 0.2835 electronvolts and the robust Huber Anchor at 0.2885 electronvolts. The Residual Bridge was worse than Identity by 0.0357 electronvolts and worse than the Huber Anchor by 0.0308 electronvolts, demonstrating how flexible nonlinear corrections can degrade when confronted with external database updates. On an entirely independent external cohort comprising 490 MC2D material groups, the Residual Bridge deteriorated further to 0.3817 electronvolts. In contrast, the frozen FidelitySwitch-2D system contained error to 0.3260 electronvolts, closely tracking the unadjusted Identity baseline at 0.3202 electronvolts and Huber Anchor at 0.3272 electronvolts. By actively steering predictions away from the failing nonlinear model, the frozen routing framework achieved a statistically verified error reduction of 0.0557 electronvolts relative to the unconstrained Residual Bridge.
Beyond point predictions, the framework integrates group-conformal uncertainty estimation to establish verifiable prediction intervals and a three-tiered triage mechanism. To prevent optimistic bias caused by multiple entries for a single material, calibration scores were calculated using the maximum normalized residual across each material group, requiring simultaneous interval coverage of every constituent row before a material group was classified as covered. Under frozen ninety percent nominal coverage parameters, observed group-simultaneous coverage reached 0.9756 across the C2DB audit and 0.9531 across the MC2D cohort, with uncertainty-error Spearman correlation coefficients registering at 0.6668 and 0.5352, respectively. Triage bins categorized candidate materials into Accept, Review, or Abstain tiers based on mean conformal interval widths. In both external cohorts, materials designated for automatic acceptance exhibited substantially heightened accuracy, achieving mean absolute errors of 0.1915 electronvolts on C2DB and 0.0744 electronvolts on MC2D, dramatically outperforming the baseline repository averages hovering near 0.3 electronvolts.
The researchers subsequently tested the transferability of the sealed framework on chemically distinct materials by evaluating a lockbox cohort of two-dimensional hybrid organic-inorganic perovskites extracted from the CoRE 2D-HOIP database. The locked evaluation comprised fifty-seven publication-disjoint material groups encompassing ninety-four rows each for band gap, conduction band minimum, and valence band maximum calculated under the r2SCAN meta-generalized gradient approximation from input Perdew-Burke-Ernzerhof values. On the primary band-gap translation task, the RiskArgminRouter delivered the lowest observed error at 0.0498 electronvolts, outperforming the uncorrected Identity error of 0.4312 electronvolts, the Huber Anchor at 0.0672 electronvolts, and the Residual Bridge at 0.0527 electronvolts. For conduction band minimum predictions, the router likewise achieved the top performance at 0.0351 electronvolts compared to 0.0577 electronvolts for Identity and 0.0469 electronvolts for Huber Anchor. Concurrently, frozen Accept subsets encompassing thirty-three band-gap groups and forty conduction-band groups concentrated exceptional accuracy, yielding mean absolute errors of 0.0187 and 0.0194 electronvolts, respectively.
To evaluate how statistical translation improvements influence downstream laboratory candidate prioritization, the authors conducted an exploratory post-open screening consequence audit across a fixed nineteen-threshold grid spanning zero to four electronvolts. The RiskArgminRouter lowered mean threshold misclassification to 0.0074, improving upon the Residual Bridge error of 0.0092 and the uncorrected Identity error of 0.0896. Crucially, the router rectified two critical misclassifications generated by the residual model at the 2.50 and 3.00 electronvolt thresholds without introducing any decision errors or harmful substitutions across the fixed grid. In fixed-window screening evaluations, the router matched the residual model in the 0.5-to-1.5 and 1.0-to-2.0 electronvolt ranges while reducing false classifications from three down to two in the broader 1.5-to-3.0 electronvolt window. Chemistry-stratified analyses revealed that the most pronounced decision benefits occurred within iodide perovskite groups, where routing slashed mean absolute error from 0.0272 to 0.0176 electronvolts and cut threshold misclassification from 0.0093 down to 0.0031.
While demonstrating that prediction-sealed routing provides measurable protection against negative transfer, the authors emphasize clear methodological boundaries and caveats regarding universal deployment. In an exploratory simulation-to-experiment benchmark translating computational calculations to experimental optical band gaps across twenty-five material groups, a standard gate achieved an error of 0.1996 electronvolts, but the sample size was considered insufficient to support broad claims of experimental translation. Furthermore, observed conformal coverage for the hybrid perovskite conduction band minimum fell to 0.8246, deviating from the nominal ninety percent target and signaling property-specific calibration strain. The study explicitly notes that the evaluated evidence remains constrained to specific computational databases, selected routes, and electronic band edges, leaving unexamined other critical properties such as work functions, effective masses, exciton binding energies, and van der Waals dispersion energies. Consequently, the research demonstrates that while simplicity-aware routing offers an auditable defense against model degradation, extensions to broader computational materials domains must remain verified by independent, pre-sealed evaluation lockboxes.
Subject of Research: Computational translation of electronic properties across two-dimensional materials databases
Article Title: Prediction-sealed risk routing limits negative transfer in cross-workflow translation of two-dimensional material properties
Article References: Talaat, A. S. (2026). Prediction-sealed risk routing limits negative transfer in cross-workflow translation of two-dimensional material properties. Scientific Reports, 16(1), Article 38557. https://doi.org/10.1038/s41598-026-75789-2
Image Credits: AI Generated
DOI: 10.1038/s41598-026-75789-2
Keywords: Two-dimensional materials, Cross-workflow translation, Density-functional theory, Negative transfer, Risk routing, Conformal prediction, Electronic band gap, Conduction band minimum, Valence band maximum, FidelitySwitch-2D, Material-group validation, Hybrid organic-inorganic perovskites, C2DB database, MC2D database, JARVIS-DFT, 2DMatPedia, Selective regression
Cite Scienmag News
Kathleen Holcomb. (November 22, 2026). Prediction-Sealed Risk Routing Limits Negative Transfer in Two-Dimensional Materials Databases. Scienmag. https://scienmag.com/prediction-sealed-risk-routing-limits-negative-transfer-in-two-dimensional-materials-databases/
Kathleen Holcomb. "Prediction-Sealed Risk Routing Limits Negative Transfer in Two-Dimensional Materials Databases." Scienmag, 22 November 2026, https://scienmag.com/prediction-sealed-risk-routing-limits-negative-transfer-in-two-dimensional-materials-databases/. Accessed 22 November 2026.
Kathleen Holcomb. "Prediction-Sealed Risk Routing Limits Negative Transfer in Two-Dimensional Materials Databases." Scienmag. November 22, 2026. https://scienmag.com/prediction-sealed-risk-routing-limits-negative-transfer-in-two-dimensional-materials-databases/








