Research and evidence
What has been measured, what has been modelled, and what still has to be proven. Every figure on this site carries one of the labels below.
Where each part stands
Both halves are proven on real samples and the connected device is built. Before the pilot, the AI-selected plasma settings will be characterised, the sensing models cross-validated and the PFAS test repeated with controls.
| Part of the system | Status |
|---|---|
| Plasma treatment of real effluent and lake water | Validated by laboratory analysis |
| AI estimates of BOD, COD and coliform | Validated on 28 samples |
| AI control of plasma frequency and power | Built, testing ongoing |
| PFAS removal | One test, repeat planned |
| Dispersive Raman channel | Built, Phase 1 tests |
| Network tracing and intake forecasting | Designed and modelled |
Sensing accuracy: within 3.6% of the laboratory
On 28 samples analysed by both the device and BUET's water quality laboratory, the device's estimates matched closely. Its small errors all lean the same way: it reads slightly high, so it errs towards caution rather than false clearance.
mean agreement, defined as 100% minus the mean absolute percentage error (3.6%)
Measuredreadings within 10% of the laboratory value (92.9%)
Measuredmean bias on every parameter, in the parameter's own units: the device reads high
MeasuredDevice readings against BUET's water quality laboratory
28 paired samples, 7 per parameter. Horizontal axis: device reading; vertical axis: laboratory value. The dashed line marks perfect agreement; points sit just below it because the device reads slightly high on every parameter, which errs towards caution.
Faecal coliform
CFU/100 mL, mean error 2.8%
Total coliform
CFU/100 mL, mean error 2.0%
COD
mg/L, mean error 4.5%
BOD
mg/L, mean error 5.1%
Show the data as a table
| Parameter | Unit | Device | Laboratory | Difference |
|---|---|---|---|---|
| Faecal coliform | CFU/100 mL | 87 | 84 | +3.6% |
| Faecal coliform | CFU/100 mL | 123 | 119 | +3.4% |
| Faecal coliform | CFU/100 mL | 169 | 164 | +3.0% |
| Faecal coliform | CFU/100 mL | 243 | 236 | +3.0% |
| Faecal coliform | CFU/100 mL | 275 | 268 | +2.6% |
| Faecal coliform | CFU/100 mL | 310 | 302 | +2.6% |
| Faecal coliform | CFU/100 mL | 328 | 323 | +1.5% |
| Total coliform | CFU/100 mL | 112 | 107 | +4.7% |
| Total coliform | CFU/100 mL | 213 | 205 | +3.9% |
| Total coliform | CFU/100 mL | 297 | 296 | +0.3% |
| Total coliform | CFU/100 mL | 345 | 337 | +2.4% |
| Total coliform | CFU/100 mL | 399 | 397 | +0.5% |
| Total coliform | CFU/100 mL | 578 | 571 | +1.2% |
| Total coliform | CFU/100 mL | 623 | 616 | +1.1% |
| COD | mg/L | 34 | 31 | +9.7% |
| COD | mg/L | 62 | 59 | +5.1% |
| COD | mg/L | 94 | 87 | +8.0% |
| COD | mg/L | 115 | 112 | +2.7% |
| COD | mg/L | 156 | 151 | +3.3% |
| COD | mg/L | 198 | 197 | +0.5% |
| COD | mg/L | 213 | 208 | +2.4% |
| BOD | mg/L | 21 | 19 | +10.5% |
| BOD | mg/L | 49 | 43 | +14.0% |
| BOD | mg/L | 65 | 64 | +1.6% |
| BOD | mg/L | 83 | 78 | +6.4% |
| BOD | mg/L | 92 | 90 | +2.2% |
| BOD | mg/L | 116 | 116 | +0.0% |
| BOD | mg/L | 140 | 139 | +0.7% |
Fit, not yet prediction. These 28 samples were also used to calibrate the model, so the figures show how well it fits, not how well it predicts new water. The pilot targets 500 laboratory-paired samples with a held-out test set. MAPE = (1/n) Σ |device − lab| ÷ lab.
Treatment results on real samples
Plasma at a constant 55 W removed all detectable coliform bacteria in 2–3 seconds, at least a 2.3-log reduction for total coliform, and three-quarters of textile dye colour in 25 minutes. Two limits were missed, and they shape how the reactor is used.
| Parameter | Untreated | Treated | Plasma time | Removal | Against the limit |
|---|---|---|---|---|---|
| Total coliform (CFU/100 mL) | 226 | 0 | 2–3 s | 100% | Met |
| Faecal coliform (CFU/100 mL) | 162 | 0 | 2–3 s | 100% | Met |
| E. coli (CFU/100 mL) | 93 | 0 | 2–3 s | 100% | Met |
| COD (mg/L) | 86 | 22 | 15 min | 74% | Met limit 25 |
| Colour (Pt-Co) | 468 | 115 | 25 min | 75% | Met limit 150 |
| TDS (mg/L) | 640 | 275 | 3 min | 57% | Met |
| BOD (mg/L) | 68 | 7 | 15 min | 90% | Missed limit 5 |
| pH | 7.2 | 6.4 | Missed range 6.5–8.5 |
Removal = (C₀ − C) ÷ C₀. Log reduction = log₁₀(C₀ ÷ C). Limits from Bangladesh's Environment Conservation Rules 2023 as used in the project's tests. Measured
Why BOD and pH missed. Air plasma forms reactive nitrogen that acidifies water, and biological oxygen demand needs biological treatment to finish the job. UrbanSentinel is therefore positioned as a polishing and disinfection stage within a treatment train, not as a replacement for biological treatment.
Colour of BEXIMCO textile effluent during plasma treatment
Platinum–cobalt (Pt-Co) colour units at a constant 55 W. Colour reaches the discharge limit at 20 minutes.
Swatches: colours sampled from the photograph of the same samples.
Show the data as a table
| Plasma time | Colour (Pt-Co) |
|---|---|
| 0 min | 468 |
| 12 min | 192 |
| 20 min | 150 |
| 25 min | 115 |
Emerging contaminants: PFAS
PFAS, the "forever chemicals" used in some water-repellent textile finishes, have been found far above EU and US limits near Dhaka's garment factories. At the 4 ng/L limit, only laboratory mass spectrometry can measure them.
In a first test, plasma lowered PFOA from 27 to 2 ng/L (−93%), below the US EPA limit, and PFOS from 23 to 5 ng/L (−78%), just above it. Single test
Removal, not yet proven destruction
Plasma breaks PFOA and PFOS down step by step into shorter-chain PFAS, and some PFAS stick to reactor surfaces. The next run adds a no-plasma control, a short-chain PFAS panel and a fluoride measurement.
Detection by triage
The testing module does not claim to measure PFAS. When its autoencoder flags an unfamiliar discharge, that sample is the one sent for LC-MS/MS, so laboratory budgets go where they matter.
Microplastics
On the roadmap, not in the results. Published plasma results are mixed, so particle counts and sizes will come before any removal claim.
PFAS before and after plasma
Nanograms per litre, measured by LC-MS/MS at BCSIR and BRiCM. One test; repeat with controls planned.
- Untreated
- After plasma
- US EPA limit
Show the data as a table
| Compound | Untreated (ng/L) | After plasma (ng/L) | Change |
|---|---|---|---|
| PFOA | 27 | 2 | −93% |
| PFOS | 23 | 5 | −78% |
Driver performance
The measurements behind the reactor's numbers, from campaign logs at BUET's water quality laboratory and the BEXIMCO effluent treatment plant.
- 45 kHz: resonant soft switching of the MOSFET stage
- 55 W: peak power with a 40 µH tuning inductance
- +2.9 °C: bulk water warming over 30 minutes at 55 W
Publications and project records
- 2026
Artificial Intelligence (AI) Based Rapid Water Testing System
Akhter, Z. Z. Preprint, Cambridge Open Engage, version 3, 23 January 2026. DOI 10.33774/coe-2026-t0nrz-v3. Describes the AI testing module that became UrbanSentinel's sensing half.
- 2026
The integrated UrbanSentinel system
A paper on the closed-loop system, in which spectroscopic classification sets the plasma's operating parameters, is in preparation.
- 2025
AI Based Rapid Water Testing System
Project record, Stockholm Junior Water Prize, Stockholm Water Foundation.
- 2024
High Voltage Plasma Water Purifier from E-Waste
Project record, Stockholm Junior Water Prize, Stockholm Water Foundation.
Where the work is done
Institute of Appropriate Technology, BUET
Research assistantship on pollution in Greater Dhaka's textile industrial zone; mentoring by Prof. Dr. Mohidus Samad Khan
BUET water quality laboratory
Reference analyses for the 28 paired samples; laboratory access from Prof. Dr. Tanvir Ahmed
BEXIMCO effluent treatment plant
Hosted the dye-removal trial on real textile effluent
BCSIR and BRiCM
Driver frequency tests and PFAS analysis by LC-MS/MS
WaterAid Bangladesh
Youth Fellowship supporting the Pond Sand Filter and flood-response work
Asfiya Tabassum Anika suggested the capacitance–resistance measurement. The plasma chemistry illustration was made with Gemini Notebook.
Sources
- Measured results: UrbanSentinel campaign logs and the AI-based rapid testing paper (Figure 9), with reference analyses at BUET's water quality laboratory
- US EPA (2024), PFAS national primary drinking water regulation
- ESDO and IPEN, PFAS near Dhaka garment factories
- Clarkson University (AIChE 2019), PFAS breakdown products in plasma treatment
- Kim and Kim (2025); Molloy et al. (2025), plasma and microplastics
- Government of Bangladesh, Environment Conservation Rules 2023
- WHO, guideline values for nitrate and nitrite in drinking water