From Soft Sensors to Smart Advisors: Agentic AI for Real-Time VFA and Alkalinity Management in Anaerobic Digestion

Introduction Volatile fatty acid (VFA) and alkalinity monitoring remain critical yet challenging aspects of anaerobic digestion operations. While VFA/alkalinity ratios below 0.3-0.4 indicate stable operation, conventional laboratory analysis introduces 24-48 hour delays that limit proactive response to process upsets. Hardware sensors capable of real-time VFA measurement require substantial capital investment and maintenance in harsh digester environments. This gap between continuous monitoring needs and practical measurement limitations has driven interest in data-driven soft sensors that estimate VFA and alkalinity from readily available process measurements. This research advances beyond conventional soft sensor approaches by integrating machine learning predictions with process modeling and natural language interfaces, creating an 'agentic' digital twin that predicts stability parameters, diagnoses root causes, recommends corrective actions, and evaluates intervention scenarios on demand. Methodological Approach The soft sensor framework was developed using four years of operational data from a full-scale co-digestion facility processing food waste, fats/oils/grease, and municipal sludge. The dataset encompassed over 150 process variables including biogas flow and composition, digester temperature, pH, mixing intensity, feed characteristics, and laboratory measurements of VFA, alkalinity, ammonia, and COD. Feature selection employed Spearman correlation analysis, mutual information-based redundancy minimization (mRMR), and gradient boosting importance ranking to identify optimal predictor subsets. Multiple machine learning architectures were evaluated including linear regression, Lasso regularization, random forest, XGBoost, and neural networks using repeated k-fold cross-validation. The selected models predict total VFA concentration, bicarbonate alkalinity, and VFA/alkalinity ratio from standard SCADA instrumentation: biogas flow rate, methane and CO2 composition, digester temperature, pH, and feed flow rates. The agentic layer integrates calibrated soft sensors with a SUMO-based process model through Python APIs, enabling real-time simulation capabilities. Natural language processing allows operators to query the system conversationally: 'Why is biogas production declining?' triggers automatic comparison against historical patterns and model predictions. 'What happens if we accept 5,000 gallons of FOG tomorrow?' executes a forward simulation incorporating the proposed feed change. The framework was subsequently deployed at a blackwater treatment pilot in Girona, Spain, where source-separated blackwater undergoes anaerobic treatment in a CSTR configuration, testing soft sensor transferability from high-strength co-digestion to lower-strength domestic waste streams. Results and Discussion The co-digestion soft sensors achieved prediction accuracy (R²) of 0.85-0.92 for VFA concentration and 0.88-0.94 for alkalinity across validation datasets spanning seasonal variations and feed composition changes. Feature importance analysis identified biogas flow rate, methane percentage, pH, and temperature as dominant predictors, collectively explaining over 75% of VFA and alkalinity variance. This confirms operators can achieve meaningful stability monitoring without dedicated VFA analyzers using only standard instrumentation. The agentic interface demonstrated particular value during feed upset events, identifying developing instability 6-12 hours before VFA/alkalinity ratios exceeded warning thresholds. Initial deployment at Girona confirmed soft sensor transferability, with recalibration requiring only 45 days of site-specific data to achieve comparable prediction accuracy. The conversational interface reduced operator response time to stability queries from 15-30 minutes of dashboard navigation to under 2 minutes of natural language interaction. Conclusions Integrating machine learning soft sensors with process models and natural language interfaces creates a fundamentally different operational tool than conventional monitoring systems. The agentic digital twin transforms anaerobic digestion management from reactive response to proactive optimization, enabling operators to anticipate stability issues and evaluate interventions before implementation. The framework's successful transfer between facilities suggests broad applicability across anaerobic systems while minimal instrumentation requirements lower adoption barriers.