A Deep Learning Framework for Global Alignment of Historical GC-MS Data Toward AI-ready Geochemical Workflows

Historical GC-MS archives contain a unique record of petroleum geochemistry, but their reuse in large-scale analytics is limited by acquisition variability accumulated over decades and by the difficulty of placing chromatograms from distinct campaigns into a common retention-time framework. These issues are especially critical for machine learning applications, which depend on consistent data representation to extract robust geological patterns. In this work, we propose a global alignment workflow designed to transform legacy GC-MS data into AI-ready geochemical datasets. The study used 18,000 chromatograms from Petrobras oil samples, focusing on the m/z 191 ion trace, which captures hopanes and related pentacyclic terpanes widely applied in petroleum characterization. The preprocessing stage comprised baseline correction with asymmetrically reweighted penalized least squares, peak extraction, and normalization by total peak area. Historical peak interpretations were then consolidated into a reference table of average retention times, enabling an initial segment-wise linear adjustment that served as supervision for model training. A one-dimensional U-Net was trained to convert non-aligned chromatograms into globally aligned signals, learning a stable mapping between raw input and reference-aligned output. Training and data preparation were performed on Petrobras internal high-performance computing infrastructure (Atena cluster). The resulting model converged with low validation loss and produced aligned chromatograms that preserved relevant peak geometry while reducing time-axis inconsistencies. Performance assessment using explained variance indicated a clear gain relative to conventional alignment strategies, and supervised UMAP projections showed better separation of samples according to geological basin, highlighting the value of the aligned dataset for downstream classification and interpretation tasks. The proposed approach provides a scalable route for standardizing historical chromatographic collections and can be extended to other analytical contexts, including additional ion traces and time-dependent techniques such as HPLC, SFC, and CE.