Experimentally Guided Neural Network and Statistical Forecasting of Membrane Water/Salt Selectivity with Minimal Mean Errors

Membrane life and performance are crucial factors in adopting membrane-based processes for water treatment and separations. This study investigated various time series models using hold-out validation of experimentally generated water vapor flux and salt rejection rates. Membrane properties were optimized by incorporating carbon-based nanomaterials to enhance anti-wetting and porosity, developing correlations between membrane characteristics and high fluxes. Fine-tuned Autoregressive Integrated Moving Average (ARIMA), Prophet, Exponential Smoothing, and Neural Prophet models were trained on an experimental dataset (N=434) collected over 36 hours to forecast performance for 72 hours. Results demonstrate the superiority of the Exponential Smoothing statistical model in predicting and forecasting membrane performance, yielding the lowest root mean square error (RMSE) of 0.006 and mean absolute error (MAE) of 0.007. This outperformance is attributed to its non-linear data fitting approach, which employs weighted averages to mitigate non-stationary behavior in time series data, a characteristic often observed in membrane performance over time. While other models showed promise, they did not match the accuracy of Exponential Smoothing in this context. The proposed modeling approach offers a more efficient alternative to traditional experimental studies, potentially leading to significant cost and time savings in the research and development phase of membrane distillation processes. This method's applicability to various membrane types and operational conditions warrants further investigation.