Machine learning-based time-series forecasting prevents electrode corrosion in organic electrochemistry

Abstract

A real-time monitoring and predictive modeling framework was developed to address electrode corrosion and process deviations in the single-step electrochemical synthesis of cortisone to adrenosterone. The approach combined inline Fourier-transform infrared (FT-IR) spectroscopy with partial least squares (PLS) modeling for rapid quantification of chemical species, alongside optical coherence tomography (OCT) for electrode surface monitoring. A novel OCT image analysis was introduced enabling efficient dynamic pixel-based quantification of electrode corrosion in real time. To complement this, machine learning-based time-series forecasting models, including convolutional neural networks (CNNs) and long short-term memory (LSTM) architectures, were optimized for long-term concentration forecast from real-time sensor data. The incorporation of domain-inspired additional features like the applied current value and the relative corrosion measure resulted in enhanced accuracy. To ensure chemical plausibility a concentration conservation constraint was employed. Despite challenges arising from fluctuations in FT-IR/PLS-derived concentration data and experimental variability, the combined monitoring–forecasting framework demonstrated robustness and predictive value. This work highlights the potential of integrating advanced sensing with machine learning-based predictive modeling for future predictive process control in electrochemical reactors.