AQUANEURON: HOW WE SIMULATED A NANOSENSOR FOR RURAL GROUNDWATER TESTING
Over 600 million Indians drink water contaminated with arsenic, fluoride, or lead — and no affordable field sensor detects all three simultaneously. We tried to change that with graphene oxide, DNA aptamers, and a Random Forest classifier.
India has a groundwater crisis that most people do not know about. The Central Ground Water Board has documented unsafe arsenic in 153 districts, unsafe fluoride in over 400 districts, and heavy metal contamination in industrial corridors nationally. Taken together, over 600 million Indians are exposed to at least one of these toxins daily.
The existing detection methods are either too expensive (ICP-MS: Rs. 2,500 per sample, results in 3-7 days) or too limited (field kits: one contaminant per kit, high false-positive rates in complex matrices). AquaNeuron is a computational study establishing the theoretical foundation for a field-deployable sensor that can detect arsenic, fluoride, and lead simultaneously at WHO-compliant sensitivity.
Why Graphene Oxide?
Graphene oxide (GO) is a two-dimensional carbon nanomaterial with ~2,630 m2/g surface area, rich surface chemistry from carboxyl and hydroxyl groups, and tunable electrical conductivity. When analytes bind to its surface, they measurably alter its electrical resistance. We model an rGO (reduced graphene oxide) surface — where mild reduction partially restores the sp2 carbon lattice while keeping carboxyl groups available for aptamer conjugation.
DNA Aptamers as the Recognition Layer
Aptamers are short single-stranded DNA sequences that fold into 3D structures binding specific targets with antibody-like affinity. We selected three SELEX-validated aptamers from published literature: arsenic (24-mer, Kd = 18.5 ppb), fluoride (16-mer, Kd = 32.1 ppb), and lead — the GBI-16 G-quadruplex (17-mer, Kd = 12.4 ppb). Aptamers are conjugated to the rGO surface via EDC-NHS chemistry, a standard bioconjugation technique. When a target analyte binds, the aptamer's conformational change alters the charge-transfer resistance of the electrode.
The Simulations
Isotherm analysis: We fit Langmuir and Freundlich adsorption models to 300 Monte Carlo bootstrap replicates. The Langmuir model dominated (R2 ≥ 0.975 vs Freundlich R2 ≤ 0.93), confirming monolayer binding — physically consistent with the monodisperse EDC-NHS conjugation. Negative delta-G values (-24.8 to -27.6 kJ/mol) confirm spontaneous thermodynamically favourable binding.
EIS simulation: Using a Randles circuit model, Rct dropped from 3,200 Ohm (no analyte) to 1,100 Ohm upon arsenic binding — consistent with published experimental data for comparable rGO-aptamer electrodes.
Monte Carlo LOD: Propagating measurement uncertainties (n=5,000) gave 95% confidence intervals: As [0.62-1.08 ppb], F [3.91-6.82 ppb], Pb [0.43-0.87 ppb]. All well below WHO limits.
The Random Forest Classifier
The edge AI layer is a Random Forest classifier (500 trees) trained on 1,500 synthetic groundwater samples across five classes: Safe, As-High, F-High, Pb-High, and Multi-Contaminated. 5-fold stratified cross-validation gave 97.3% accuracy with AUC > 0.999 for all classes. The RF architecture was chosen over neural networks because it runs in under 2 seconds on an Arduino Nano 33 BLE Sense after m2cgen C++ conversion — neural networks would exceed the 256 KB RAM limit. Interpretability via MDI feature importance also matters for regulatory acceptance in public health applications.
What is Next
Phase 2 is physical fabrication at a NABL-accredited laboratory, pending funding. The key unknowns are real-world performance in complex geological matrices and sensor stability in tropical Indian field conditions. All code is open-source at github.com/prateektiwariii/AquaNeuron.