Chetanya Goyal supervised by Dr. Aftab M. Hussain  received his – Master of Science Dual Degree in Electronics and Communication Engineering (ECD). Here’s a summary of her research work on Machine Learning Enhanced Colorimetric Analyte Measurement

Water quality monitoring is critical for public and industrial water supplies. Current methods present a duality. Test strip and visual comparison methods are low-cost but suffer from run-to-run variabil- ity, whereas spectrophotometric methods are accurate, but costly and require trained operators to ana- lyze results. This work presents a fully automated, low-cost colorimetric system employing a Machine Learning (ML) inference pipeline. The system integrates four primary hardware subsystems: (i) a novel cereal-dispenser-based automatic powder reagent dispenser fabricated in Acrylonitrile Styrene Acrylate (ASA) via Fused Deposition Modelling (FDM) 3D printing; (ii) an automatic, voltage-controlled Peri- staltic Pump liquid reagent dispenser capable of drop accurate dosage, with a peak flow rate of 10.8 mL/min; (iii) a 3D printed Acrylonitrile Butadiene Styrene (ABS) sample chamber (50 mm ×50 mm x40 mm external dimensions) with a surrounding fluid handling system to automate cleaning cycles; (iv) an optically-baffled Red-Green-Blue (RGB) Light-Emitting Diode (LED)-Light-Dependent Resis- tor (LDR) sensor pair that measures dark, unreacted sample, and white references along with reacted sample color values as an RGB triplet. The measured color values, along with other supplementary features, are used as the feature set for the downstream ML pipeline. Electrical connections are consol- idated onto a Printed Circuit Board (PCB) and components are housed in an IP-67 dust and water-proof enclosure. A Flask-based Hypertext Transfer Protocol (HTTP) server with a browser-based frontend exposes Application Programming Interface (API) endpoints to either the individual hardware compo-nents or to run assays. A dataset of 180 color-concentration pairs, spanning the 0-10 Parts Per Million (ppm) free-chlorine range, was created in 10 day-wise batches to train and evaluate multiple super-vised ML Regression models. Initial evaluation with a train-test split and reference-based normalization yielded an R2 of 0.980, a Root Mean Squared Error (RMSE) of 0.39 ppm, and a Mean Squared Log Error (MSLE) of 0.021 for an optimised Stacking Regressor. Shapiro-Wilk and Breusch-Pagan sta- tistical tests confirmed non-gaussianity and heteroscedasticity, motivating the use of a more rigorous Leave-One-Batch-Out (LOBO) evaluation methodology. The deployment-ready configuration uses an RGB triplet normalized with a RobustScaler and an optimised Random Forest model, achieving an RMSE of 0.848 ppm, an R2 of 0.901, and an MSLE of 0.0184. This configuration was obtained after conducting normalization, feature, and model ablation as well as hyperparameter tuning. The presented platform is designed to be analyte-agnostic, and colorimetric reagents can be refilled and hot-swapped without modifying the sensor or ML inference subsystems. By eliminating manual reagent dosing, visual color matching, and operator intervention, the system provides a modular framework for real-time, scalable, and repeatable water quality monitoring.

June 2026