How Does a Chemistry PDF Become Data an AI Can Use?

OpenDataLab at Shanghai AI Laboratory released the MinerU.Chem technical report with collaborators from East China Normal, Tongji, Fudan, Shanghai Jiao Tong, Peking University and other institutions. It asks a concrete question: how can an organic-chemistry PDF become data that can be searched, computed over, audited and reused by models?^[S1]

General PDF parsers can already identify titles, paragraphs, formulas and tables. Chemistry introduces another layer. A reaction scheme contains substrates, products, catalysts, temperatures and sequence for a chemist, yet may remain a single rectangular image block to a document system.^[S1]

Finishing PDF parsing does not mean the chemistry inside has been parsed. MinerU.Chem adds the missing layer after document parsing: chemistry parsing.

Not one model, but a pipeline

MinerU.Chem does not ask one model to jump from a whole PDF page to a reaction database. MinerU first parses the page; a chemistry-specific filter then decides whether a figure or table contains chemistry, keeping microscopy and ordinary charts away from more expensive specialist models.^[S1]

The system then locates molecular drawings and nearby identifiers such as 1a, 2b or compound 7, linking visual structures to references in the paper. Only then does OCSR reconstruct a molecular graph and export machine-readable formats such as MolFile and SMILES.^[S1]

Reaction-diagram parsing comes last: which molecules are reactants and products, whether text near an arrow is reagent, solvent or temperature, and how several arrows form successive steps. Together, the five modules turn image recognition into a document-to-data pipeline.^[S1][S3]

A SMILES string without knowing whether the paper calls it 2b or 7c is hard to audit. A set of molecules without reaction relationships is still a stack of molecule cards, not a synthesis route.

Why SMILES alone is not enough

Real papers contain more than textbook SMILES: Ph, Et, Bu, Markush fragments, polymer repeat units, wavy bonds, coordination bonds and unusual valence. If the figure says Ph but extraction expands it into a full phenyl ring to produce a standard SMILES, the output may be chemically valid while no longer faithfully matching the source.^[S2][S1]

CARBON preserves not only atoms and bonds but also 2D coordinates, abbreviations and complex properties that ordinary SMILES struggle to carry, then exports common formats from that richer intermediate representation. The principle matters more than the format name: first preserve what the paper actually shows; standardization comes second.^[S2][S1]

The more interesting question is not how many points AI OCR gained, but where AI-for-chemistry training data will come from. Decades of synthesis knowledge live in papers, supporting information and patents, often in reaction schemes and molecular drawings. Multimodal models can look at images, but as long as the knowledge stays in images it is hard to feed directly into workflows that need structured molecule and reaction records, which is what reaction prediction, retrosynthesis, condition modeling and training data for new chemistry foundation models all require.

MinerU.Chem aims to shorten the distance from human-readable literature to machine-usable scientific data. The hard part is not merely running models at scale, but preserving extraction and provenance together.

A Molecule Summary List stores source location, compound identifier, structure representation and MolFile/SMILES for each molecule. A Reaction Summary List then organizes reactants, products and conditions and links them back to molecule records.^[S1]

The part worth noticing is not JSON itself, but traceability. If stereochemistry is wrong, a reviewer can return to the page and original bounding box. A fluent summary is hard to debug; a structured record with source location can be sampled, relabeled and corrected.^[S1]

A useful way to read MinerU.Chem is as a chemical evidence index layered over a PDF.

The easiest number to misread is 93.02%. It is not document-level extraction accuracy, nor a score over all 5,024 structure images, but the SMILES exact match of one structure-recognition module on the 2,392 samples that convert to valid SMILES. Detection and identifier matching come before it; reaction parsing and document-level linking come after it. End-to-end correctness is a different measure.^[S1]

On the same benchmark the report also ran three kinds of comparison system: specialist OCSR tools and chemistry-tuned vision-language models, commercial chemical OCR services, and general multimodal models. The structure module leads all of them; the next best SMILES exact match is GPT-5.6-Sol at 74.87%, 18.15 points behind. The report then splits the benchmark by its MOSAIC difficulty labels into three mutually exclusive subsets of rising visual and chemical-semantic complexity: 98.28% SMILES / 92.15% graph accuracy on the easiest, 93.49% / 80.11% on the visually harder one, and 70.13% / 55.42% on the subset carrying chemical-semantic difficulty labels. The graph-accuracy margin over the next best system is printed as 43.25 points in the introduction and 42.1 in the results, so it is not quoted here.^[S1]

The two metrics have different denominators. SMILES exact match can only be scored on samples with a valid SMILES annotation: 2,392 overall, and 1,219, 875 and 298 in A, B and C. Structures with, among other things, Markush fragments, unconventional abbreviations or bonds that break valence rules cannot become valid SMILES and are left out. Graph accuracy is scored on all 5,024 images: 1,987, 1,976 and 1,061 in A, B and C. The three subsets do not overlap and together make up the whole benchmark. The versions differ too: the first MolRecBench-Wild paper reports 5,029 structures from 820 papers; MinerU.Chem uses the latest revised version, 5,024 images all with graph annotations; 2,392 is the subset of those convertible to SMILES.^[S1][S2]

This is a technical report and deployed system, but its strongest quantitative results concern individual modules rather than end-to-end document extraction. A mismatched compound identifier or condition attached to the neighboring arrow can still yield a perfectly valid-looking but incorrect database record.^[S1]

Complex Markush structures, R-groups, cross-page references and multi-step routes remain clear boundaries. MinerU.Chem is better understood as high-throughput first-pass processing and a doorway to human review. The critical metrics are how many correct records a paper yields and how much human time is needed to repair the rest.^[S1]

Scientific document parsing begins where OCR ends

A chemistry reaction scheme is a graph; a materials paper may contain a phase diagram; a biology paper, a pathway. A general parser finds what is on the page. A domain system must understand what it means scientifically.

Scientific data needs provenance, not only accuracy

An incorrect stereocenter silently entering a training database is different from a chatbot making a verbal mistake. Source page, bounding box and compound ID may not look like AI, but they are core infrastructure.

The most useful automation is often not unattended

Large-scale machine extraction that focuses human time on difficult Markush structures, low-confidence cases and cross-page linking may be more valuable than insisting on full autonomy.

The literature remains a data mine not yet fully available to AI

As models improve, architecture is not always the bottleneck. A great deal of valuable scientific knowledge may simply remain unstructured.