Skip to content

Back to the catalog

Repairing PBE-Spurious Metallicity for HSE06-Level Screening of 2D Photocatalysts for Green Hydrogen Production
arxiv.orgpaper

Repairing PBE-Spurious Metallicity for HSE06-Level Screening of 2D Photocatalysts for Green Hydrogen Production

This paper addresses a failure mode in semilocal PBE DFT calculations that mislabels narrow-gap 2D semiconductors as metals, causing viable photocatalysts to be excluded from screening in the Computational 2D Materials Database (C2DB). The authors build a two-stage ML pipeline: a classifier to repair PBE-spurious metallicity using structural/chemical/magnetic descriptors, and a curated XGBoost regressor for HSE06-PBE delta-learning that reconstructs HSE06 band gaps with MAE of 0.108 eV (R-squared=0.989) versus 1.036 eV for raw PBE, validated against targeted VASP HSE06 calculations. The corrected screen yields 28 retained relaxed candidate green-hydrogen photocatalysts, several of which would have been wrongly excluded at the PBE level.

TL;DR XGBoost delta-learning corrects PBE-spurious metal misclassification to enable HSE06-accurate screening of 2D green-hydrogen photocatalysts.
Open original ↗
Document

What it claims

The classifier half of this pipeline is weak, and the paper is explicit about it: the labeled true-metal class holds only 29 materials, so Stage I is used only for triage, not decision-making. Its best holdout gives 87.5% accuracy but 0.286 true-metal recall and 0.643 balanced accuracy — it misses most real metals. The regression half is the strong claim (MAE 0.108 eV, R²=0.989 vs 1.036 eV raw PBE), but the page's actual thesis is methodological: high-fidelity correction "must be evaluated by candidate membership, not only by global regression error." Low aggregate MAE does not guarantee correct screening decisions.

How it works

Two stages over C2DB, with deliberate leakage control. Stage I classifies HSE06-unknown PBE metals from structural, chemical, magnetic, and stability descriptors while excluding HSE06/GW quantities and PBE electronic shortcuts (features that would leak the answer). Stage II is a Δ-learning regressor over the corrected insulating population, predicting E_g^HSE − E_g^PBE.

flowchart TD
  A[C2DB PBE results] --> B[Stage I: triage PBE metals]
  B -->|corrected insulators| C[Stage II: Δ-learn HSE-PBE gap]
  C --> D[pH 0 screen, 1.6-2.8 eV window]
  D --> E[VASP HSE06 check on 6 compounds]
  E --> F[28 retained relaxed candidates]

Evidence

The page compares three feature sets against targeted VASP HSE06 on six ML-predicted compounds, giving material-level MAEs of 0.417 eV (C2DB-native), 0.182 eV (Magpie+structural), and 0.110 eV (curated) — the curated model wins at the compound level, not just globally. The gap window is 1.6–2.8 eV at pH 0. The corrected screen yields 10 strict and 29 initial relaxed candidates; 4 strict and 18 relaxed would fail the same window at PBE. One compound (1AgBr-1) shifts above the upper cutoff under VASP verification, dropping the relaxed count from 29 to 28.

Limits

  • Stage I is not a reliable metal detector (0.286 recall on 29 labeled metals); treat it as a filter that needs human/DFT follow-up, not an oracle.
  • VASP verification covers only six compounds — small confirmation set behind the retained-candidate count.
  • Abstract-only page: feature lists, dataset sizes for training, and per-candidate identities beyond 1AgBr-1 are not given here.

Organisations named here

  • C2DB
  • Computational 2D Materials Database

Technologies named here

  • VASP
  • XGBoost

Techniques named here

  • Leakage-aware repair of PBE-spurious metallicity
  • HSE06–PBE Δ-learning

Related in research