Optimization of Anaerobic Digestion through In-situ Hydrogen Biomethanation and Machine Learning Prediction of Methane Production

Abstract

Biogas produced through anaerobic digestion (AD) represents an important renewable energy carrier, yet the relatively low methane (CH4) content of raw biogas limits its value as fuel and grid-injectable gas. In-situ hydrogen (H2) biomethanation (HBM) the direct injection of H2 into an operating anaerobic digester to promote hydrogenotrophic conversion of carbon dioxide (CO2) into additional CH4 offers a promising route for biological biogas upgrading. However, the low aqueous solubility of H2 means that gas-liquid mass transfer is frequently the rate-limiting step, and existing Anaerobic Digestion Model No. 1 (ADM1) implementations do not adequately represent the dynamic interplay between H2 supply, transfer intensity, and process stability. This thesis presents a hybrid mechanistic, data-driven framework to optimize in-situ HBM and predict CH4 production using machine learning. The standard ADM1 model, implemented within the Benchmark Simulation Model No. 2 (BSM2) structure using the open-source PyADM1 platform, was extended in three ways: (i) an external H2 injection term was incorporated into the gas-phase H2 balance; (ii) gas-specific volumetric mass transfer coefficients (k_L a) for H2, CO2, and CH4 were introduced using diffusivity-based scaling relationships; and (iii) an inorganic carbon limitation factor was added to the hydrogenotrophic methanogenesis rate to prevent unrealistic H2 conversion under carbon-depleted conditions. The modified model was verified against published biomethanation benchmark data, achieving deviations below 8% for CH4 production and below 1% for pH across the stable operating range. Long-term dynamic simulations were conducted over 280-day periods across a matrix of k_L a scenarios (60 to 1000 d-1) and H2 injection rates (0 to 2500 m3.d-1), yielding a dataset of approximately 280 daily observations per scenario. Results showed that CH4 production increased consistently with H2 injection under all k_L a conditions, but the efficiency of H2 utilization depended strongly on mass transfer intensity. Under low k_L a conditions (60 d-1), H2 conversion efficiency reached only 57-61%, with substantial H2 slip and declining CH4 content at higher injection rates. Under high k_L a conditions (1000 d-1), H2 conversion efficiency exceeded 96% across the full injection range, with CH4 content in the outlet gas reaching up to 84%. An injection-efficiency evaluation framework was developed to identify the most favorable operating points under each mass-transfer scenario for subsequent machine-learning (ML) dataset preparation. Four ML algorithms Support Vector Regression (SVR), Random Forest (RF), Extreme Learning Machine (ELM), and Evolutionary Polynomial Regression (EPR) were trained using ten process variables to predict CH4 and total biogas flow rates. SVR achieved the highest testing accuracy, with R2 values of 0.996 for CH4 and 0.9948 for biogas, followed by ELM with 0.9899 and 0.9834, respectively. RF also demonstrated strong predictive performance, with testing R2 values of 0.9879 for CH4 and 0.9807 for biogas. EPR produced lower but still strong predictive performance, with testing R2 values of 0.9561 and 0.9276, while offering the added advantage of explicit, interpretable mathematical expressions. The findings demonstrate that in-situ H2 biomethanation can substantially improve CH4 yield and biogas quality when gas-liquid mass transfer is sufficiently strong, and that the combined ADM1 ML framework developed in this work provides both process insight and practical prediction tools for anaerobic digestion optimization. The EPR-derived explicit formula further provides a transparent and computationally inexpensive alternative for CH4 estimation within the investigated operating domain.