RESEARCH AREAS
At the forefront of the Fourth Industrial Revolution, CAPS Lab is redefining the boundaries of Process Systems Engineering. Our philosophy is simple yet transformative: the massive data streams generated by modern industrial and natural systems hold the key to solving humanity's most complex challenges. By integrating Machine Learning, Control Theory, and Optimization, we build autonomous systems capable of reasoning, learning, and adapting. From optimizing biopharma supply chains to predicting climate dynamics, our work bridges the gap between theoretical AI and real-world impact.
1. Decision AI: Intelligent Control & Optimization
Real-world processes are often too complex or uncertain for traditional mathematical models. We leverage Reinforcement Learning to bridge this gap, creating agents that learn optimal control strategies directly from data and adapt to plant-model mismatches in real-time.
Beyond control, we are pioneering RL for Combinatorial Optimization. We tackle "NP-hard" challenges—such as dynamic sensor network design and cutting plane selection—using novel Hierarchical and Multi-Agent RL frameworks. By developing dynamic meta-agents, we are building solvers that outperform traditional integer programming methods in both speed and adaptability.
2. AI for Sustainability: Agriculture & The Circular Economy
Sustainability is no longer just a goal; it is an operational constraint. Our research expands the system boundary beyond the factory floor to the ecosystem level. We are pioneering AI-driven frameworks for the Food-Energy-Water-Land-Fertilizer (FEWLF) Nexus, using multi-objective optimization to guide crop land allocation and resource usage under future climatic scenarios.
In parallel, we are driving the transition to a Circular Economy. We utilize Natural Language Processing (NLP) and Large Language Models (LLMs)—such as our Recycle-BERT and CCU-Llama architectures—to mine vast scientific literature. This allows us to automatically extract Life Cycle Inventory (LCI) data, accelerating the discovery of sustainable technologies for plastic recycling and carbon capture.
3. Generative AI: Large Models for Scientific Discovery
We are entering the age of "AI-driven Science." Our lab is at the forefront of developing domain-specific Large Language Models (LLMs) that decode the language of chemistry and engineering. We have developed a suite of specialized models—including CCU-Llama (for Carbon Capture & Utilization), Sustain-Llama, and Battery-Llama—that automatically extract complex synthesis recipes and property data from millions of scientific papers.
By converting unstructured text into structured knowledge, these generative models accelerate the discovery of sustainable materials and energy storage solutions, reducing the timeline for innovation from years to weeks.
4. Industrial AI: Predictive Maintenance & Fault Diagnosis
In the era of Industry 4.0, a system must do more than operate; it must self-diagnose. We are moving beyond traditional statistical monitoring to Deep Generative and Graph-based approaches. Our lab develops advanced algorithms—including Graph Neural Differential Equations (GNDAE) and Generative Adversarial Networks (GANs)—to detect early-stage anomalies in complex, non-linear chemical processes.
We focus on "interpretable AI," ensuring that our models not only detect faults but explain their root causes. Recent work also explores Machine Unlearning to adapt models dynamically to changing process states without costly retraining, ensuring 24/7 reliability in safety-critical environments.
5. Climate AI: Atmospheric Modeling & Prediction
The Indian economy and safety are intrinsically linked to the monsoons. We are revolutionizing meteorological forecasting by fusing Physics-Informed Machine Learning with traditional General Circulation Models (GCMs). Our research targets the Sub-seasonal to Seasonal (S2S) prediction gap, offering more accurate forecasts for the Indian Summer Monsoon and tropical cyclones.
By employing Deep Learning architectures like LSTMs and Deep Neural Networks, we mine historical atmospheric data to uncover latent patterns that traditional physics models might miss. This hybrid approach improves the granularity of rainfall predictions and helps mitigate the risks associated with extreme weather events like floods and droughts.
6. Foundational AI: Graph Learning & Knowledge Mining
Many real-world systems—from molecular structures to transportation networks—are best represented as graphs. Our fundamental research in Graph Machine Learning focuses on modeling the time-series dynamics of these networked systems. We are developing "frugal" AI models through Graph Distillation techniques (like Bonsai and Mirage) that deliver high performance with low computational overhead.
Additionally, we are advancing Scientific Knowledge Extraction. By integrating Graph Neural Networks with Large Language Models, we are building systems capable of reading, understanding, and structuring millions of scientific papers to automate the discovery of new materials and chemical processes.
