<p>Accurate estimation of unbound gas and water volumes is essential for evaluating gas reserves and optimizing well placement in complex reservoir systems. Traditional multi-mineral (MM) analysis requires predefined physical property matrices and computationally expensive iterative forward modeling, which fails when mineral count exceeds available logs. This study presents machine learning (ML) and deep learning (DL) frameworks that eliminate these requirements by directly learning nonlinear relationships between well-logs and volumetric contents. The workflow was applied to predict mineral volumes (anhydrite, calcite, dolomite, illite) and fluid volumes (unbound gas and water) in a complex Iranian carbonate gas reservoir using wireline log data (sonic, gamma ray, neutron porosity, bulk density) from ten wells. Four ML models were developed and optimized using Bayesian optimization: k-nearest neighbors (kNN), adaptive boosting (AdaBoost), random forest (RF), and extreme gradient boosting (XGBoost). A multi-channel neural network (MC-NN) architecture was then proposed, integrating outputs from the four optimized models with raw well-log inputs. The MC-NN achieved enhanced predictive accuracy, with R² = 0.94 and MAE = 0.0007 for water content prediction on the test set, compared to individual models (XGBoost: R² = 0.90, MAE = 0.0008; RF: R² = 0.80, MAE = 0.0012). Model uncertainties were quantified through five-fold cross-validation (R² standard deviation: 0.012–0.028), with prediction-level confidence intervals of ± 0.01–0.02. The proposed methodology eliminates the physical properties matrix and iterative computations required in traditional MM analysis, enabling instant predictions even in underdetermined systems. This data-driven approach offers a robust framework for estimating mineralogical and fluid compositions in carbonate reservoirs and contributes to more informed reservoir characterization and production optimization.</p>