Wastewater Dilution Test — 50056

Experiment Overview

On 20 May 2026, Lume sensor 50056 was tested in a controlled wastewater dilution series at the Water Quality Lab in the Civil Engineering Department at San Diego State University, across eight concentration steps: 0%, 1%, 5%, 10%, 25%, 50%, 75%, and 100% raw wastewater in tap water. Each window lasted approximately 8–13 minutes. Two reference sensors were run in parallel: an Aqualog benchtop EEM fluorometer (IFE-corrected, treated as ground truth) and a Turner C3 in-situ fluorometer measuring CDOM (365 nm) and TLF (280 nm). Lab turbidity (NTU) was measured by grab sample, and grab samples at four concentrations were processed by IDEXX Colilert for E. coli MPN.

The LUME sensor swept LED power (64–1024) and SiPM bias (2800–3400 V) continuously throughout the experiment, generating 108,173 raw readings across 44 LED×bias combinations. The calibration combo (LED 512 / bias 3000) was used for fluorescence analysis; TOF backscatter (SPS, kcps/SPAD) provided an independent turbidity channel.

Top-Level Findings

1. Inner filter effect (IFE) is the dominant constraint on raw fluorescence above ~25% wastewater. Every Lume LED×bias combo and the Turner C3 TLF channel peak around 25–50% and decline afterward — the limit is the sample's UV optical density (OD ≈ 0.05 at 280 nm), not detector saturation.
2. Layering temperature + turbidity corrections substantially recovers linearity for both sensors. Bedell 2022 thermal compensation (Lume ρ = −0.03/°C; Turner factory cal −0.00576/°C) followed by Skinner 2024 turbidity recovery (k = −0.004/NTU) lifts the single-predictor calibration fit substantially:
       Lume: R² 0.681 (raw) → 0.804 (corrected), inferred-WW% RMSE 24.3 pp → 17.5 pp (MAE 20.9 → 14.7 pp).
       Turner C3: R² 0.397 (raw) → 0.613 (corrected), inferred-WW% RMSE 43.6 pp → 28.2 pp (MAE 38.8 → 25.3 pp).
   The temperature step alone adds essentially nothing on either sensor (this experiment varied only ~2 °C); the turbidity step does all the heavy lifting.
3. Lume outperforms the Turner C3 across the dynamic range — hardware-level, and the gap survives after corrections. At the 0%→1% low-end step the Lume TLF resolves the change at SNR = 4.73 vs Turner C3 TLF's 2.70 — ~75% more reliably per reading. In the linear regime (0–25% WW) the two TLF channels are statistically equivalent (raw R² 0.921 vs 0.908). In the IFE regime (50–100%) the Turner TLF reverses direction (raw R² = 0.289) while the Lume's signal compresses more gracefully (R² = 0.798) — ~3× more predictive where conventional in-situ fluorometers fail. After both sensors are run through the same correction chain, the Lume retains its lead: Lume corrected R² = 0.804 (WW% RMSE 17.5 pp) vs Turner corrected R² = 0.613 (WW% RMSE 28.2 pp).
4. Lume has a unique full-range channel that Turner lacks: SPS (TOF backscatter). SPS tracks the Aqualog truth at R² = 0.999 across all 8 windows, including the 50–100% IFE regime where every fluorescence channel (including Turner CDOM) degrades. In high-strength wastewater, particle loading is a more robust proxy for organic concentration than fluorescence — and the Lume gets it from the same instrument, on the same timestamp, with no extra sensor.
5. Colilert grab samples carry ~1,000× inferential uncertainty about raw E. coli concentration — even though the underlying physics is a clean linear dilution. Because every step is the same wastewater sample diluted in sterile DI water, MPN must fall exactly linearly with WW%; any departure is measurement error, not biology. Yet single-dilution extrapolations to 100% range from 12,000 to 13,000,000 MPN/100 mL depending on which dilution is chosen as the anchor, and the 10%-WW sample fell three orders of magnitude below what the 100% sample predicts. This is a direct measurement of how variable Colilert (and by extension, any single-grab microbial test) is on a sample whose true concentration is known to be linearly related to the dilution — the failure mode that motivates continuous in-situ sensing without dilution choices or 24-hour incubation.

How to use this page: click any series in the chart legends to show/hide that trace, and use the checkboxes in the Concentration Annotations table near the bottom to include or exclude individual experimental windows — all charts and regressions re-fit instantly to the active selection, so you can stress-test any finding by holding out a window.

TLF Signal / Turbidity / Temperature — Sensor 50056

Turner C3 — In-Situ CDOM, TLF & Temperature

Turner C3 grab-sample time series collected in parallel with the dilution experiment. Timestamps are on the Turner’s own clock (not aligned with LUME UTC). Each concentration window has 180 readings at ~1 Hz; background shading shows the annotated windows. Note CDOM continues rising through 75–100% while raw TLF (tryptophan-like) peaks at ~25% and decreases at high concentrations — inner filter effect on the 280 nm excitation channel. Three TLF traces are plotted: light-blue raw, mid-blue temp-corrected applying the factory thermal-compensation formula TLF/(1 + ρ·(T−20)) with ρ = −0.00576/°C, and burnt-orange temp + turbidity corrected additionally multiplying by exp(−k·NTUlab) with k = −0.004/NTU (Skinner 2024 / Saraceno 2017 IFE recovery, lab grab-sample NTU per window). The temperature correction is small (~1–2 °C across the experiment); the turbidity correction lifts the signal at 50–100% where IFE attenuates raw TLF.

Colilert E. coli Reference — Dilution Response and Lume Comparison

Grab samples at four concentrations were diluted in sterile DI water (no tap water) and processed by IDEXX Colilert (Quanti-Tray/2000) for E. coli MPN/100 mL; the 0% blank read 0 (non-detect, < 1 MPN/100 mL). The left panel below tests internal consistency: because every step is a simple dilution of one wastewater sample, MPN should fall linearly with WW%. Each dotted line is a two-point regression through the 0% blank and one other dilution — the raw-WW E. coli concentration you would infer if you trusted only that single step. Solid teal line is the OLS fit through all four points. The right panel compares the Lume TLF signal to the matched Colilert reading on a semi-log axis (Colilert spans ~4 orders of magnitude).

Findings: The three single-dilution extrapolations span 12,000 – 13,000,000 MPN/100 mL at 100% — a ~1,083× range that depends entirely on which dilution is chosen as the anchor. The 50% and 100% lines agree within ~25% (10.3 M vs 13.0 M projected raw-WW), but the 10% extrapolation lands more than three orders of magnitude below the direct 100% reading — a 1,200 MPN/100 mL measurement implies only ~12,000 MPN/100 mL at 100%, vs the 12,997,000 actually observed.

The all-points OLS line has R² = 0.986, which looks excellent — but the 100% point's magnitude dominates the total variance, so R² is not a useful goodness-of-fit metric here; the 10% residual is a thousand-fold off the line yet contributes almost nothing in absolute terms. Because the Colilert dilutions were prepared in sterile DI water, chlorine carryover is not a candidate explanation for the 10% discrepancy. Remaining plausible causes include osmotic / viability loss in the 90%-DI sample before reagent mixing (DI water is hypotonic and depletes intracellular electrolytes), Poisson MPN imprecision at low well counts, or a serial-dilution bookkeeping error in the lab. The right-hand panel shows the same story from the sensor side: the Lume TLF response tracks log(Colilert) closely across 50% and 100%, but the 10% Colilert reading sits well off the relationship the other three points define. Regardless of cause, the operational takeaway is that a single Colilert grab sample carries ~3 orders of magnitude of inferential uncertainty about the underlying microbial load — the failure mode that motivates continuous in-situ sensing without dilution choices or 24-hour incubation.

Lume Correction Ladder — How Each Correction Step Recovers Linearity

Lume cal-combo TLF (mon2) plotted against the experimentally-set WW% (x-axis, independent variable) at each correction stage, with an OLS calibration fit (dashed teal). The caption beneath each panel reports R², the regression slope (counts per percent), the intercept (signal at WW% = 0), the in-signal-units residual error (RMSE, MAE, max |residual|), and the equivalent inferred WW% error from inverting the calibration. Read the improvement off the captions: a good correction lifts R², shrinks residuals, and tightens inferred-WW% error.

Turner C3 Correction Ladder — Same Analysis, Reference Sensor

The same correction-ladder analysis applied to the Turner C3 TLF channel for comparison. Temperature correction uses the factory ρ = −0.00576/°C with the Turner's own temperature; turbidity correction uses the same k = −0.004/NTU as the Lume ladder, with the lab grab-sample NTU per window as input. Putting both sensors through the same correction pipeline lets you read off whether the corrections help Turner as much as they help the Lume, and whether the Turner's hardware-level fit is comparable across the dynamic range.

Inner filter effect (IFE): At high concentrations, the sample absorbs its own excitation and emitted light, causing the fluorescence signal to plateau then decrease. The IFE onset is a property of the sample's optical density, not the detector — changing LED power scales all signal values proportionally but does not shift the concentration at which the peak occurs (the peak is at OD ≈ 0.05 regardless of I₀). Implication for this dataset: if 50% and 75% are already indistinguishable at LED 512 (flat top of the IFE curve), they will remain indistinguishable at any other LED power because the curve shape is unchanged — only its amplitude shifts. What the comparison below can reveal is whether any combo suffers from additional SiPM microcell saturation on top of IFE (signal plateaus earlier than expected), vs. combos where the SiPM is in its linear range and the true IFE shape is visible. Lines marked ⚠ are non-monotonic (IFE-dominated). Use Normalize to compare curve shapes independent of LED power. One combo per LED level shown (bias nearest to cal).

Add Annotation

Concentration Annotations