Interpretable generative machine learning platform for chronological prediction of forensic paper aging via GC-IMS...

The determination of paper aging is crucial for forensic document examination, enabling the identification of temporal origin in forgery or fraud cases. Paper aging is difficult to distinguish due to the high compositional complexity, intricate temporal evolution mechanisms, and challenges in tracing diagnostic markers, limiting reliable chronological authentication. Herein, an interpretable generative end-to-end machine learning platform integrating gas chromatography-ion mobility spectrometry (GC-IMS) with machine learning was established for prediction of the temporal origin of forensic paper aging. Sixty-three differential volatiles were identified in paper materials via GC-IMS analysis. Chronological classification and regression machine learning models were constructed based the volatile fingerprints from papers aged 1 to 35 years. The CatBoost model achieved the best overall performance in both age-class identification and year prediction based on characteristic volatile profiles. SHAP (SHapley Additive exPlanations) model identified key markers as the most influential predictors, such as 3‑hydroxy-2-butanone-D, Hydroxyacetone, Cyclopentanone, and Isomenthone, enabling paper samples specific attribution of prediction outcomes. An end-to-end chronological prediction software tool was developed to automate data import, age prediction, and interpretation result generation, thereby improving its practicality and accessibility for forensic applications. Overall, this study provides a deployable analytical framework for interpretable chronological prediction of forensic paper aging and is expected to generate reliable results for judicial authentication.