Abstract
Conventional methods for determining shale oil content are costly and time-consuming when large numbers of samples must be analyzed. This study developed and evaluated a general regression neural network (GRNN) to predict shale oil content using a compiled global dataset comprising 94 oil-shale observations. Following predictor-redundancy analysis, eight variables were retained as model inputs: analytical moisture, ash, total sulfur, and the elemental composition of kerogen (C, H, S, N, and O). The selected NeuralTools (NT) GRNN yielded R2 = 0.950, RMSE = 2.02 percentage points, and MAE = 1.30 percentage points for the 75 training observations. For the 19-observation hold-out testing subset, the corresponding values were R2 = 0.871, RMSE = 2.35 percentage points, and MAE = 1.95 percentage points. Repeated 10-fold cross-validation using an independently implemented standard GRNN produced a pooled out-of-fold (R2) of 0.595, RMSE of 5.47 percentage points, and MAE of 3.81 percentage points, demonstrating sensitivity to data partitioning. Under identical repeated cross-validation partitions, multiple linear regression provided lower mean RMSE and MAE than the independently implemented GRNN. Ash was the dominant predictor in the selected NT model. The results support the use of the GRNN as a preliminary screening and sample-prioritization tool but also demonstrate the importance of robust validation when modeling relatively small and heterogeneous geological datasets.