Multi-fidelity optimization of hybrid biofilter media using physics-informed neural networks and Mordukhovich variational analysis

Abstract

Biofilters are widely used to remove volatile organic compounds (VOCs) from industrial exhaust streams, yet their real-world performance often deteriorates before the end of the planned service period due to media aging, moisture imbalance, and increasing pressure drop. Most existing design and optimization approaches treat the filter bed as static and focus on short-term removal efficiency, while ignoring (i) time-dependent degradation, (ii) the practical trade-off between efficiency, longevity, and operational burden, and (iii) uncertainty and limited experimental data that make optimization recommendations fragile when operating conditions change. As a result, solutions that look optimal in a snapshot can lead to rapid decline and frequent maintenance in long-term operation. This paper proposes an adaptive, multi-objective optimization framework for long-term biofilter media design and operation. The method learns how removal performance evolves over time and uses this predictive capability to search for implementable media ”recipes” and operating setpoints that jointly maximize removal efficiency and stability while limiting pressure drop and experimental cost. The framework is validated on three physically built modular biofilters (135 observations per configuration; 405 total) treating acetone as a representative oxygenated VOC, including a three-stage design with a zeolite buffering layer. The best optimized configuration achieved high initial and sustained performance (initial removal up to 94.7% and median removal up to 88.89%) with an extended operational lifetime of up to 301 days, while exhibiting slower performance deterioration with less negative degradation slopes (from -0.0160% day-1 to -0.00916% day-1) compared to conventional media selection and standard optimizers.