CEOE PhD Student Seminar Series: 1 (of 3)

ABSTRACT

Speaker 1 (Md Rahadujjaman): Predicting specific disinfection byproduct (DBP) species, rather than broad regulatory categories, is crucial for assessing drinking water toxicity. Most Quantitative Structure-Activity Relationship (QSAR)-based machine learning (ML) models rely on generic descriptors that ignore halogenation mechanisms or depend solely on treatment conditions. We introduce an explainable QSAR-based ML framework that predicts species-specific DBP formation from phenolic precursors by integrating structural descriptors with operational conditions. We created a Curated Molecular and Quantum Descriptor Set (CMQDS) comprising 34 molecular and 11 quantum-chemical descriptors representing phenol halogenation mechanism. Using study-grouped splitting to prevent literature data leakage, our Gradient Boosting Regression Trees (GBRT) model with CMQDS outperformed general-purpose libraries (e.g., Mordred, RDKit, Morgan fingerprints), achieving a test R2 of 0.729 compared to 0.510 for the best alternative. It exhibited minimal overfitting (test-train gap of 0.15) and generalized to an external dataset (R2 = 0.614). SHapley Additive exPlanations (SHAP) interpretation confirmed the reliance on established chemistry, such as meta-dihydroxyl ring cleavage and acid-base speciation. Ablation analysis revealed that using only reaction conditions decreased external R2 from 0.614 to 0.191, whereas relying solely on molecular structure yielded a negative R2 and performed worse than predicting the mean. The k-nearest neighbors (KNN)- based applicability domain (AD) was defined and applied to the European Chemicals Agency (ECHA) REACH inventory, covering at least 4,000 phenolic substances. The model is available as an open-access web app to predict species-specific DBPs while reporting the query precursor's domain status to elucidate predictive reliability.

Speaker 2 (Moheb Henein): Operational camera networks, such as the U.S. Geological Survey (USGS) Hydrologic Imagery Visualization and Information System (HIVIS), provide critical real-time visual observations for hydrological and atmospheric monitoring. However, real-world outdoor deployments face persistent challenges from adverse weather, atmospheric degradation, and complex co-occurring environmental conditions. This seminar presents a unified computer vision framework designed to automate image degradation detection, restoration, and atmospheric understanding for large-scale camera networks.

First, we introduce a two-stage framework for real-world image dehazing featuring a high-precision classifier (99.28% accuracy) paired with a novel, one-step diffusion-based generative model (CycleGAN-Turbo) to filter and restore degraded river imagery. Second, we present CLIMAT (CLIP-based Localized Intercorrelated Multilabel Atmospheric Transformer), a vision-language model leveraging region-aware prompt tuning and transformer-based fusion to capture spatially distributed weather cues and multi-label interdependencies across outdoor scenes. Finally, we explore ongoing efforts to utilize Large Language Models (LLMs) for automated dataset annotation of degradation and weather conditions, establishing the foundation for a unified, all-in-one image restoration system built specifically for real-world environmental data.

BIOGRAPHY

Md Rahadujjaman is a first-year Environmental Engineering PhD student in the Department of Civil, Environmental and Ocean Engineering at Stevens Institute of Technology at Stevens Institute of Technology, advised by Dr. Tao Ye. His research applies interpretable machine learning to understand and predict the formation of disinfection byproducts in drinking water.

Moheb is a third-year Ph.D. student in Civil Engineering, advised by Prof. Marouane Temimi. His research applies computer vision techniques to detect and enhance degraded images from a nationwide river-monitoring camera network. Additionally, he develops downstream applications to assess hydrological conditions, including river ice dynamics and water surface velocimetry.