Executive Summary
Even in the age of scientific AI, the task that truly eats time before training begins is confirming that the data is usable — and the standard for "usable" changes from one field to the next. Data that perfectly honors the FAIR principles (Findable, Accessible, Interoperable, Reusable) can still be useless for AI training if it lacks structure or if its representations don't line up. This article uses REDI, an automated data-readiness framework from U.S. national labs, to show how a single set of names — completeness, consistency, fitness — gets translated into entirely different rules in each domain.
REDI turns the massive datasets of four scientific domains — climate, proteomics, materials, and fusion — into AI-training-ready data automatically, running on supercomputers. Fusion simulation data goes through a normalization that compresses the original by 298×; protein data has to place its gap token at exactly index 21 or training breaks silently. The same pipeline is used across all four, yet the point of intervention differs by domain. And REDI itself admits that the deeply domain-specific transformations still resist full automation.
When the domain-specific diagnosis Pebblous DataClinic has done for image data extends to science data at national-supercomputer scale, what carries over unchanged and what has to be built anew? Two things shift: the scale of the compute, and who renders the verdict on validation — where the human eye gives way to the laws of physics.
298×
Fusion data compression
Normalizing XGC1 electron density to [0,1] shrinks it 298× vs. the original
Index 21
Protein gap-token position
Put it at 20 by mistake and training corrupts silently, with no error
86%
Ingest share of fusion runtime
The bottleneck moves by domain — climate spends 70% on the output stage
Levels 3–5
Where each domain stands
Materials and climate near level 5; fusion and life sciences sit at 3–4
'AI-Ready' Is a Different Word in Every Field
The flashier model architectures get, the easier it is to forget one fact. In a scientific AI project, the work that actually consumes the most time before training starts is confirming that the data is in usable shape and converting it into that shape. Because it's tedious, this step often gets skipped or rushed — but data that goes wrong here comes back as poor performance no matter how good a model you stack on top.
A paper from researchers in the U.S. national laboratory system, "Automated Data Readiness for Scientific AI," confronts one misconception head-on: honoring the FAIR principles does not make data ready for AI. FAIR data can still lack structure, carry representations that contradict one another, miss the metadata needed for feature extraction, or fail to match the format training demands. The paper calls this broader idea "Data Readiness for AI" — a notion of data quality that spans schema consistency, feature completeness, provenance traceability, governance constraints, and validation against domain expectations.
This article's central claim goes one step further. The data-quality criteria carry the same names in every field: completeness, consistency, fitness. But the rules those names actually point to differ completely by domain. In climate data, "completeness" means the grid is aligned with no gaps; in genomic data it means personal information has been properly anonymized; in fusion data it means the laws of physics are preserved within a precision tolerance. Same word, different yardstick.
REDI, a National-Lab Automated Checking Pipeline
The framework the paper proposes is called REDI. It's a five-stage pipeline that automatically converts massive scientific datasets into AI-training-ready data, following the sequence ingest → preprocess → transform → structure → output, with provenance instrumentation built into each stage for reproduction and distribution. A companion tool, SetGo, automates FAIR compliance and catalog publication.
The scale is not laboratory-sized. REDI achieved near-ideal parallel scaling out to 100 nodes on a Frontier-class supercomputer. The goal is not raw performance but something else: turning data preparation — until now locked inside individual researchers' scripts — into a reproducible, reusable community asset. It's an attempt to let the know-how of preparing data be shared and verified the way papers and code are.
The heart of REDI is that the stage names stay identical across the four domains while what actually happens inside each stage is completely different from one domain to the next — a structure that pairs a domain-neutral skeleton with domain-specific parameters. That difference comes into focus in the measured numbers of the next section.
Same Name, Different Yardstick — Four Domains Measured
The four domains REDI validated are climate science (ClimaX), proteomics (OpenFold), materials science (HydraGNN), and fusion (XGC1). Setting alongside them the genomics case from a conceptual predecessor paper (Brewer et al., AI Magazine 2026) — Enformer and DNA sequence — sharpens the contrast. The table below summarizes the actual transformation that turns "raw data" into "AI-ready data" in each domain, and what the result is validated against.
| Domain (case) | Core transformation | Domain-specific validation criterion |
|---|---|---|
| Climate (ClimaX) | Regrid a non-uniform 0.94°×1.25° mesh to a uniform grid + z-score normalization | Grid alignment and correlation — Pearson r ≥ 0.9999 |
| Genomics (Enformer, DNA) | One-hot encode DNA sequence + tile splitting, anonymization | PHI/PII regulatory compliance and format integrity |
| Proteomics (OpenFold) | Deepen MSA (5,907 → 8,919 sequences) + sequence encoding | Place the gap token at exactly index 21 — misplace it and training corrupts silently |
| Materials (HydraGNN) | Normalize energy to per-atom values (-4.76 eV/atom) + build a 6Å-radius graph | Energy-conservation precision — Pearson r = 1.000, filter unphysical structures |
| Fusion (XGC1) | Min-max normalize electron density to [0,1] (298× compression) + 3D mesh projection | Physical-quantity reconstruction precision — RMSE = 0.000 |
Reading the second column top to bottom shows how a single criterion — completeness — splits into strikingly different things. In climate, completeness is a question of whether any coordinate is missing from the Earth's surface grid. In genomics, whether personal information remains is the first hurdle, before missing values even come up: data that fails to clear regulation cannot be put into training no matter how clean it is. In fusion, the question is whether the physical relationships in electron density survive normalization — checked to decimal-point precision.
The proteomics gap-token case captures, in miniature, why domain-specific validation resists automation. In a multiple sequence alignment, the gap token that marks an empty position has to sit at index 21. Counting through the 20 amino acids, it's easy to absent-mindedly place it at 20 — and if you do, the model learns wrong sequences without a single error message. It's a trap no general-purpose data-validation rule will ever catch, one that only someone who knows protein biology would know.
Even the Bottleneck Sits Somewhere Different by Domain
Domain specificity doesn't end at the validation criteria. Which stage the time cost of data preparation piles up in also differs by field. By REDI's instrumentation, fusion spends 86% of its runtime in the ingest stage that reads the data in, and materials spends 73% in that same ingest stage. Climate, by contrast, concentrates 70% of its time in the output stage that stores results. The answer to "where should I optimize?" points somewhere different in each domain. Because much of this bottleneck comes from data format and I/O methods, the paper notes that — before touching any algorithm — format choice becomes the first lever for per-domain optimization.
This observation dovetails with the picture drawn by REDI's conceptual predecessor. Brewer et al. divided data readiness into five levels: a scale that begins at Raw, moves through Cleaned, Labeled, and Feature-engineered, and arrives at Fully AI-ready. Apply the same scale, and each domain sits at a different point on it right now.
Climate and materials are already close to levels 4–5, because their pipeline patterns are well established and benchmark datasets are plentiful. Fusion, by contrast, stays at levels 3–4: its data is sparse, noisy, and access-restricted. Life sciences, including genomics, sits around level 3, blocked by the walls of regulatory compliance and format mismatch. To borrow the paper's phrasing, "readiness is a spectrum, not a dichotomy." The moment you ask AI-ready or not as a yes/no question, the question is already wrong.
What DataClinic Already Knew — and What It Has to Learn
This idea is not foreign to Pebblous. DataClinic has already diagnosed 134 image datasets — 12 million images — designing five signals (integrity, balance, pixel diversity, feature distribution, and separability) and mapping them to the ISO 5259 standard. It built diagnostic items tailored to the domain of images. What REDI is trying to do in climate, fusion, and genomics rests on the same idea: that AI-ready cannot be settled by a single general-purpose checklist.
So when an approach that worked for image data extends to science data at national-supercomputer scale, what stays the same and what changes? What stays is the principle — that diagnostic items have to be designed anew for each domain. Two things change. One is scale: where an enterprise-grade dataset runs on a single server or in the cloud, science data demands a 100-node parallel supercomputer. The other is who — or what — renders the verdict on validation.
With image data, human visual judgment enters the signal, as it does with pixel diversity. With science data, the laws of physics themselves become the validation criterion — physical and biological constraints like energy conservation, precision tolerances, and the exact index of a gap token. What decides which data is correct is not the human eye but natural law. That is the essential difference.
This article pairs with an earlier one, "AI Now Grades AI Readiness." Where that piece dealt with who grades readiness (a shift from human checklists to AI agents), this one asks the question one step earlier: that the scorecard itself has to differ by domain. Read together, the two trace a movement from the evolution of the grader to the domain differentiation of the grading criteria.
One last thing, to be honest about it. REDI itself states that it cannot yet automatically handle the deeply domain-specific transformation sequences. Automation has produced clear gains in consistency, provenance, and reproducibility, but it does not replace the knowledge that the gap token belongs at index 21. Embedding domain specificity into the structure, then, still leaves a seat for the domain expert. For data-preparation automation to truly scale, the field has to keep answering one question: how do you translate domain knowledge into code?
References
- 1.Wilkinson, S. R. et al. (2026). "Automated Data Readiness for Scientific AI." arXiv:2607.02771.
- 2.Brewer, W. et al. (2026). "Data Readiness Pipeline Patterns for Scientific AI at Scale." AI Magazine, Wiley. DOI 10.1002/aaai.70056.
- 3.Hiniduma, K. et al. (2025). "Data Readiness for AI: A 360-Degree Survey." ACM Computing Surveys.