Neural Network-Based Tensor Completion: Advancing Predictions of Activity Coefficients and Beyond

Abstract

Although existing tensor completion methods have progressed in predicting two-and three-dimensional data, they still struggle to capture nonlinearities and temporal dependencies in relational data effectively. We introduce an innovative solution to this research gap: our novel 3D-DMF-H method for tensor completion. Developed as a neural network-based matrix completion approach, our method extends the Deep Matrix Factorization (DMF) method, handling nonlinear data structures and effortlessly incorporating additional data points. Our method applies to a wide range of three-dimensional tensor completion problems, and exceptional accuracy was achieved in predicting activity coefficients to model phase equilibria. Notably, the 3D-DMF-H method outperforms the benchmark thermodynamic g E model UNIFAC (standard) and significantly enhances the accuracy of azeotrope predictions when integrated with the equation of state PC-SAFT. Our findings demonstrate the promising potential of machine learning for advancing applications in the chemical industry and highlight the necessity for further algorithmic refinement and exploration.