A modification of the integrated dry route (IDR) process for uranium conversion in nuclear fuel fabrication to render it responsive to react and adapt to changes in measured variables is described. Data conditioning techniques and feature extraction methodologies have been developed using real-world industrial datasets comprising 1685 valid IDR process batches. Bidirectional Long Short-Term Memory (Bi-LSTM) sequence classification networks were designed, trained, and tested by framing the IDR process and its assessed fluorine content results as a classification problem. Five comprehensive experiments were conducted using features extracted from both raw process data and domain knowledge provided by experienced process operators. The results of the five training scenarios are presented using confusion matrices, which report the mean Specificity for predicting defective batches and the mean Sensitivity for predicting satisfactory batches. Additionally, a Receiver Operating Characteristic (ROC) curve is included, showing the Area Under the Curve (AUC) values for the classification outcomes in each of the four training tasks. Both the confusion matrices and the ROC curve indicate that the best-performing model is trained with a combination of raw data features and post-processed features derived from prior experimental and technical knowledge. This model achieved classification accuracies of 97% or higher, confirming that a purely data-driven approach is insufficient. The primary objective is to predict the quality of the uranium dioxide ( UO 2 ) output, specifically its fluorine content, more quickly than is currently achieved in the factory. Test case results demonstrate the effectiveness of the trained Bi-LSTM network, suggesting its potential utility in developing a digital twin for the IDR system. With further development, such models could enable on-line feedback to kiln conditions, allowing for real-time responsiveness during IDR operations.