“…Our study has breakthroughs at two distinct levels. First, at the mechanistic level, a new pathway of AD data acquisition was explored to capture physiochemical patterns/variations accurately in a real-time in situ mode via the deployment of a novel solid-state mm-sized sensor array (MEA) technology capable of monitoring multiple liquid-phase AD parameters including pH, temperature, oxidation–reduction potential (ORP), conductivity, and ammonium (NH 4 + ). ,,, For the parameters (e.g., volatile fatty acid, alkalinity) unable to measure in a real-time in situ mode, we adopted ADM1 as the soft sensor for data generation with full physiochemical backbones compared to previous data-driven soft-sensor methods. , Second, at the data level, the MLA interpretability was enhanced with the integrated data acquisition technology consisting of robust MEA Physical Sensors (MAPS) and ADM1-based Soft Sensor (ADSS), in which different ML models were analyzed with posthoc interpretation methods to extract physical meanings via the assessment of feature/AD parameter correlations and lab AD experiments (Figure a–c). With the close loop of the MAPS-MLA-ADSS-Experiment strategy, a feasible path was elucidated to conquer the long-standing interpretability problem in MLA and complexity problem in ADM1 by first “pouring” physical meanings into MLA features through MAPS and ADSS, then “brew” those meanings out of the MLA prediction results.…”