Trapped at Tier 1: Why machine learning-guided electrocatalyst discovery has not closed the lab-to-industry gap

Machine learning has transformed electrocatalyst discovery, enabling screening across millions of candidate compositions. Yet publications on HER/OER have grown by over 63% since 2014, while no comparable benchmarking tracks whether this effort has translated into measurable gains in commercial alkaline water electrolyzer (AWE) performance. This Perspective argues the disconnect is structural: every ML-guided workflow surveyed here validates at Tier 1 (rotating disk electrode, ≤10 mA cm⁻², 25°C), while industrial AWE operates at 200–1200 mA cm⁻², 80–90°C, and 30 wt% KOH — conditions characteristic of Tier 3 (AWE/PEMWE stack) — separated from Tier 1 by four phenomena absent from all ML training datasets: bubble-induced mechanical stress, dissolution under dynamic load cycling, binder–catalyst interface degradation, and high-current mass transport limitations. Tier 2 (flow cell/beaker cell at ≥200 mA cm⁻², 60–80°C, industrial electrolyte) sits between these extremes and is equally absent from published ML training pipelines. No closed-loop workflow surveyed here has completed retraining using Tier 2 or Tier 3 stability data. We define this three-tier benchmarking framework — Tier 1 (RDE), Tier 2 (flow cell), Tier 3 (stack) — identify the structural data gaps, and propose a tier-informed ML architecture with concrete recommendations for data infrastructure.