What question did this study set out to answer?

April 7, 20260 citationsOpen Access

Automated Digitalization and Molecular Enrichment of Industrial Chemical Documentation: A French-Language NLP and Cheminformatics Pipeline

AAAbderaouf Haithem Aroua AHAymen HAMIDI MBMerouane BOUKANDOURA

Key Points

The aim is to automate the conversion of French-language industrial chemical documentation into machine-learning-ready molecular datasets.
Developed an automated pipeline for processing chemical documentation.
Extracted ingredient tables and specifications using regex from OCR text.
Normalized ingredient names via a domain-specific synonym map and databases.
Enriched molecular data using the PubChem PUG REST API.
Computed molecular descriptors for unique molecules.
Achieved 100% normalization coverage of ingredient names.
Recovered canonical SMILES for 98.1% of chem compounds.
Generated 208 RDKit and 973 Mordred molecular descriptors.
Validated reproducibility on independent product data with 80.5% INCI coverage.

Abstract

Industrial chemical manufacturers maintain extensive technical documentation — product specifications, raw material datasheets, and formulation records — that encode decades of domain knowledge yet remain inaccessible to modern machine learning methods when authored in local languages as scanned or unstructured digital files. We present an automated pipeline for converting French-language industrial chemical documentation into ML-ready molecular datasets. Applied to 809 pages of technical documentation from a North African industrial chemical manufacturer (ENAD SHYMECA, Algeria) — producing detergents, cosmetics, and automotive chemical products — the pipeline processes 382 raw extracted strings — of which 333 are real ingredients (the remaining 49 are noise: water entries, metadata fragments, and mixture compositions) — through four stages: (1) regex-based extraction of ingredient tables and physicochemical specifications from Markdown-converted OCR text; (2) cascaded INCI normalization of the 333 real ingredients via a 322-entry domain-specific synonym map (Stage 0), resolving 315 names, exact matching against the EU CosIng database (24,094 entries, Stage A), and fuzzy token-sorted matching (threshold 82%, Stage B), achieving 100% normalization coverage without requiring LLM assistance; (3) molecular enrichment via the PubChem PUG REST API, recovering canonical SMILES for 255 of 260 chemically definable compounds (98.1%; 5 unresolvable entries are legitimate polymeric surfactants and pigment mixtures); and (4) molecular descriptor computation yielding 208 RDKit and 973 Mordred 2D descriptors for 99 unique molecules after quality filtering (NaN threshold 20%, median imputation to 100% completeness). Pipeline reproducibility is validated on 53 Open Beauty Facts products (1,518 ingredient instances), achieving 80.5% INCI coverage. The pipeline code and a 322-entry French-to-INCI synonym dictionary (Table S1) are released as open-source software to enable replication by other French-language manufacturers. Formula concentration data are not released due to confidentiality obligations with the industrial partner.

AIに質問

Bookmark

View Full Paper