Rae, MitchellMitchellRaeRanade, VivekVivekRanadeWalker, GavinGavinWalkerHeinrich, StefanStefanHeinrichRamachandran, RohitRohitRamachandranSingh, MehakpreetMehakpreetSingh2025-10-082025-10-082025-09-14Archives of computational methods in engineering (in Press): (2025)https://hdl.handle.net/11420/57834In twin-screw granulation (TSG), the mean residence time (MRT) of materials significantly influences granule properties, such as size distribution and density, impacting the quality of the final product. Accurately estimating MRT is crucial because deviations can lead to overwetting, compaction issues, or insufficient granulation. This study presents a hybrid approach that combines machine learning and data interpolation techniques to model MRT as a function of process parameters, including feed flow rate, screw speed, screw configuration, and liquid-to-solid ratio. Our goal is to develop a predictive tool capable of handling coarse datasets for precise MRT estimation. By optimising the MRT, process control, efficiency, and batch-to-batch consistency can be improved, ensuring adherence to product specifications and facilitating cost-effective scale-up. This study explores the integration of various univariate and multivariate spline interpolation techniques with the nonlinear autoregressive with exogenous inputs (NARX) and multilayer perceptron (MLP) machine learning methods to enhance the accuracy of MRT. While numerous studies have utilised large datasets, this study examines a coarse dataset, applying various interpolation techniques to enhance data resolution and consequently improve the performance of the NARX machine learning model. This study examined training and testing datasets of different sizes, demonstrating the versatility and applicability of the coupled methodology. Our findings demonstrate the advantages of multivariate cubic spline interpolation with the NARX approach over MLP and Kriging with univariate interpolation methods. This paper presents a comprehensive review of existing interpolation techniques and their impact on modeling performance, addressing a critical gap in the current literature. The results show that the multivariate cubic spline interpolation with the NARX approach achieved a 72% reduction in the root mean square error (RMSE) and an 85% increase in the adjusted compared to the existing Kriging interpolation technique (Ismail et al. in Powder Technol 343:568–577, 2019). In terms of computational efficiency, the NARX approach with univariate and multivariate spline interpolations are 16 times more efficient than the Kriging interpolation technique.en1886-17842025Springerhttps://creativecommons.org/licenses/by/4.0/Technology::660: Chemistry; Chemical Engineering::660.6: BiotechnologyNatural Sciences and Mathematics::519: Applied Mathematics, ProbabilitiesJournal Articlehttps://doi.org/10.15480/882.1595910.1007/s11831-025-10371-z10.15480/882.15959Journal Article