Sustainable development demands process-systems solutions that are simultaneously profitable, environmentally responsible, and socially beneficial. Meeting this triple bottom line becomes increasingly difficult as industrial systems grow in complexity and scale. When traditional optimization methods struggle under high dimensionality and large constraint spaces, purely data-driven models—and, in some cases, hybrid approaches that embed mechanistic insight—provide the flexibility and scalability needed to capture real-world process behavior. My doctoral work illustrates how AI-driven digital twins, classification-based implicit constraint handling, and adaptive optimization workflows can effectively tackle real-world industrial challenges while advancing sustainability goals across three critical domains: economic efficiency, environmental protection, and clean energy innovation.
Resource-Efficient Oil Recovery Through Intelligent Optimization
Waterflooding operations consume 250 million barrels of water daily to extract 80 million barrels of oil, creating a $40 billion annual water management challenge. These operations are further constrained by thousands of limitations—flowrate bounds, platform capacity, and water-cut thresholds—that make traditional optimization computationally intractable.
To address this, we developed a constrained reinforcement learning (RL) framework that learns optimal well-pressure strategies while adapting to evolving feasible regions. The workflow begins with offline training of two surrogate models: a deep neural network that predicts economic performance and a binary classifier that screens feasibility across all constraints [1]. The RL agent explores the action space by adjusting bottom-hole pressures, receiving rewards for predicted NPV and penalties for constraint violations, enabling a gradient-guided search toward feasible, high-value solutions. When the agent encounters low-confidence regions, the full simulator is selectively invoked to validate new points, which then incrementally trains the classifier and refines the feasibility model over time. This hybrid learning system significantly reduces simulation reliance while dynamically converging on strategies that balance economic returns with water conservation.
While optimizing existing industrial processes addresses economic sustainability, protecting public health from environmental hazards requires equally sophisticated approaches.
Rapid Chemical Risk Assessment for Environmental Emergencies
When environmental disasters release complex chemical mixtures requiring rapid evaluation, exhaustive experimental screening is often impractical, so we developed a machine learning framework that transforms high-content microscopy data from engineered biosensors into actionable insights. Rigorous preprocessing, feature engineering, and principal component analysis are used to denoise data and highlight key biological patterns. Within this framework, we trained support vector machines, random forests, and neural networks as separate classifiers, achieving up to 98% accuracy in predicting estrogen receptor activity. This approach reduces evaluation timelines from months to hours, enabling timely protection of vulnerable populations during chemical exposure events [2,3].
Beyond protecting human health from environmental hazards, addressing the root causes of environmental degradation requires innovative approaches to carbon emissions and clean energy production.
Optimizing Carbon Utilization through Physics-Informed Modeling
With atmospheric CO₂ levels surpassing 420 ppm, scalable solutions that reduce emissions and convert CO₂ into valuable chemical feedstocks are urgently needed. The reverse water-gas shift (RWGS) reaction offers one such path, converting CO₂ and renewable hydrogen into carbon monoxide (CO), a key precursor for fuels and chemicals. Membrane-assisted reactors enhance this reaction by continuously removing water, but optimizing their performance involves solving complex differential equations, which becomes computationally expensive across varied conditions. To address this, we developed a physics-informed Deep Operator Network (DeepONet) that embeds governing laws into a neural architecture, learning the nonlinear mapping from input conditions to species flowrates. Unlike conventional physics-informed neural networks (PINNs), DeepONet generalizes across scenarios without retraining. Once integrated into an optimization framework, the model rapidly identifies operating conditions that maximize CO₂ conversion and reactor efficiency—delivering scalable, high-impact solutions for carbon utilization [4].
Ready for Industry Impact
The methodological consistency across these projects—spanning data science, AI, and optimization—reflects a systematic approach to translating academic innovation into industrial value. Combined with hands-on data science internship experience in the chemical and energy sectors, this research portfolio directly addresses the computational and sustainability challenges facing chemical, energy, pharmaceutical, tech, and consulting companies. I am particularly interested in roles in R&D, data science, or innovation leadership where I can apply these methodologies to solve complex engineering problems and advance sustainable technology development. These validated frameworks are ready for immediate deployment to accelerate digital transformation and deliver measurable sustainability outcomes across process-intensive industries demanding both technical depth and real-world impact.
References
[1] Aghayev, Z., Voulanas, D., Gildin, E. and Beykal, B., 2025. Surrogate-Assisted optimization of highly constrained oil recovery processes using Classification-Based constraint modeling. Industrial & Engineering Chemistry Research, 64(15), pp.7751-7766.
[2] Aghayev, Z., Szafran, A.T., Tran, A., Ganesh, H.S., Stossi, F., Zhou, L., Mancini, M.A., Pistikopoulos, E.N. and Beykal, B., 2023. Machine learning methods for endocrine disrupting potential identification based on single-cell data. Chemical engineering science, 281, p.119086.
[3] Aghayev, Z., Walker, G.F., Iseri, F., Ali, M., Szafran, A.T., Stossi, F., Mancini, M.A., Pistikopoulos, E.N. and Beykal, B., 2023. Binary Classification of the Endocrine Disrupting Chemicals by Artificial Neural Networks. In Computer Aided Chemical Engineering (Vol. 52, pp. 2631-2636). Elsevier.
[4] Aghayev, Z., Li, Z., Patrascu, M. and Beykal, B., 2025, Advancing Sustainable CO2 Utilization with Parametric PINNs in Membrane Reactor Design. In Computer Aided Chemical Engineering (Accepted).