Integrating Petrographic Analysis with AI for Paleoenvironment Reconstruction

Thesis Abstract

This study addresses the persistent challenge of integrating qualitative petrographic observations with quantitative AI-based modeling to reconstruct paleoenvironmental conditions from sedimentary records. The central aim is to develop a hybrid workflow that combines petrographic analysis with machine learning to enhance the accuracy, reproducibility, and interpretability of paleoenvironment reconstructions in fluvial and deltaic systems. Specific objectives include (1) compiling a geo-referenced petrographic dataset from 180 hand-specimen thin sections and 60 scanned SEM-EDS images from Cretaceous deltaic deposits in a well-exposed sedimentary basin; (2) extracting quantitative petrographic features (framework grain content, mineralogical indices, textural metrics, pore fabric parameters) and coupling them with qualitative descriptors via a standardized coding framework; (3) developing and validating a supervised machine learning model ensemble (random forest, gradient boosting, and support vector regression) to predict paleoenvironment proxies such as water depth, energy regimes, and salinity; (4) testing model performance against established proxies derived from ichnology, grain-size distribution, and carbonate microfacies; and (5) evaluating model interpretability through SHAP (SHapley Additive exPlanations) analysis and a theoretical framework grounded in the neomorphic theory of petrographic indicator reliability. The methodology adopts a mixed-methods research design combining epistemic triangulation of petrographic data with data-driven modeling. The population comprises sedimentary rock samples from two Late Cretaceous deltaic successions and one contemporaneous riverine deposit, selected to maximize diversity in sedimentary facies. A stratified random sampling approach yields 180 representative thin sections and 60 SEM-EDS datasets. Data collection employs conventional petrographic microscopy for modal analyses, point-count statistics, and grain-size measurements, complemented by automated image analysis to quantify mineralogical textures and pore networks. Instrumental data include X-ray diffraction (XRD) for bulk mineralogy, thin-section cathodoluminescence for diagenetic overprint, and SEM-EDS for microtextural and geochemical context. The machine learning component uses a feature set comprising petrographic metrics (e.g., quartz K-feldspar ratio, clay mineral fraction, framework grain content), textural descriptors (asymmetry, roundness, sorting metrics), and geochemical indicators (SiO2/Al2O3, CaO/MgO). The response variables are independently derived paleoenvironment proxies obtained from established proxies including grain-flow indicators, ichnofacies assemblages, and palaeosalinity estimates. Model performance is assessed via cross-validated RMSE and R^2, with hyperparameters tuned through grid search. Model interpretability is addressed with SHAP values to identify feature contributions and to evaluate the stability of key petrographic indicators across facies. The anticipated findings indicate that integrated petrographic-AI models will outperform traditional regression approaches in predicting paleoenvironment proxies, with improved accuracy in delineating energy conditions and salinity regimes. It is expected that mineralogical indicators such as quartz content, feldspar proportion, and clay mineralogy, when combined with textural metrics, will exert significant predictive influence, while diagenetic overprint will be partially mitigated by incorporating SEM-EDS and cathodoluminescence data. The study will also reveal context-dependent feature importance across facies, demonstrating that deltaic versus fluvial settings exhibit distinct petrographic signatures in relation to paleoenvironmental parameters. A crucial contribution is the demonstration of a transparent hybrid framework that integrates petrographic expertise with machine learning, supported by SHAP-based interpretability and grounded in the neomorphic and facies association theories, thereby enhancing reproducibility and transferability of paleoenvironment reconstructions. The study contributes to knowledge by establishing a replicable, scalable workflow for combining qualitative petrography with quantitative AI modeling to reconstruct paleoenvironmental conditions, offering a template applicable to diverse sedimentary basins globally. It is recommended that future research extend the dataset to marine shelf deposits and incorporate time-series modeling to track environmental shifts through sequence stratigraphy, while refining explainable AI methods to further reduce interpretive uncertainty in petrographic-based paleoenvironmental reconstructions.