TÜBİTAK 2209-A · PeakTransformer

HQ & DA Real-Time DPV Analytics

Simultaneous quantification of Hydroquinone and Dopamine from 219-point DPV signals via a 1D-CNN + Transformer hybrid model with curriculum-learned signal decomposition.

Hydroquinone
HQ
E₀ = −0.16 V
Dopamine
DA
E₀ = +0.12 V
TARGET R²
0.99
Peak area (µM)
Mixture HQ clean DA clean
Noise Types tick → enable · sliders update live
Select Experimental Data
Compare Signals
Legend:
█ Synthetic
█ Experimental
⚡ Model Prediction
Mixture Signal
True Clean Components (ground truth — synthetic data)

Curriculum Learning Dataset

Each stage represents a level in the curriculum learning hierarchy, where the model is trained on progressively harder synthetic examples.

Phase 1 — Generic Peak Detector Phase 2 — HQ+DA Specific
1
Stage 1 — Single Peak, Clean Gaussian
Single symmetric peak • Noise-free • Model learns the basic amplitude–E₀ relationship
▸ Generation Logic
Samples: \(E_0\sim\mathcal{U}(-0.54,+0.54)\) V, \(A\sim\mathcal{U}(2,28)\) µA
Gaussian peak:
\[I(E)=A\exp\!\left[-\tfrac{1}{2}\!\left(\tfrac{E-E_0}{\sigma}\right)^{\!2}\right]\]
EMG (\(\tau\!\to\!0\approx\) pure Gaussian):
\[I(E)=\tfrac{A\sigma}{\tau}e^{\frac{\sigma^2}{2\tau^2}-\frac{\Delta E}{\tau}}\operatorname{erfc}\!\left(\tfrac{\sigma}{\tau\sqrt{2}}-\tfrac{\Delta E}{\sigma\sqrt{2}}\right)\]
\(\tau\in[0.001,0.006]\) V · single peak · noise-free
2
Stage 2 — 1–3 Peaks + AWGN
Multiple peaks • Mild white noise • Model freely learns peak count
▸ Generation Logic
\(N\in\{1,2,3\}\) independent peaks, min \(|E_{0i}-E_{0j}|>0.10\) V
AWGN noise:
\[I_\text{noisy}=I_\text{clean}+\mathcal{N}(0,\sigma_n)\] \[\sigma_n=\sqrt{\frac{P_\text{sig}}{10^{SNR/10}}}\]
All peaks share the same \(\sigma\).
3
Stage 3 — Overlapping EMG Peaks + Structural Noise
2–5 asymmetric EMG peaks • Intentional overlap • Fouling, pH shift, matrix effect
▸ Generation Logic
EMG, \(\tau\in[0.008,0.025]\) V · min sep. \(0.04\) V (intentional overlap!)
\[I(E)=\tfrac{A\sigma}{\tau}e^{\frac{\sigma^2}{2\tau^2}-\frac{\Delta E}{\tau}}\operatorname{erfc}\!\left(\tfrac{\sigma}{\tau\sqrt{2}}-\tfrac{\Delta E}{\sigma\sqrt{2}}\right)\]
4 structural noise types (\(p=0.55\), each indep.):
\(I \mathrel{\times}= e^{-k\,i}\)  (fouling)
\(\text{roll}(I,\;50\text{mV}{\cdot}\Delta\text{pH})\)  (pH shift)
\(I \mathrel{+}= aE^2+bE+c\)  (matrix effect)
+ AWGN: \(\text{SNR}\sim\mathcal{U}(25,45)\) dB
PHASE 2 — HQ+DA SPECIFIC — 7-STAGE v5 CURRICULUM
4
Stage 4 — HQ+DA Clean EMG (Baseline)
Noise-free • Langmuir amplitude • Model learns the concentration–area relationship • 8K samples
▸ Noise Profile (stage "4")
All noise sources disabled — EMG peaks only.
Langmuir isotherm:
\[A=I_{\max}\cdot\frac{K_\text{ads}\cdot C}{1+K_\text{ads}\cdot C}\]
AWGN=0 • Baseline=0 • pH=0 • Fouling=0 • Ionic=0
4a
Stage 4a — AWGN
White noise only • First robustness to noise • 5K samples
▸ Noise Profile (stage "4a")
\[I_\text{noisy}=I_\text{clean}+\mathcal{N}(0,\sigma_n)\] \[\sigma_n=\sigma_\text{std}\cdot I_\text{MAX},\quad\sigma_\text{std}\sim\mathcal{U}(0.01,0.05)\]
AWGN=1.0 • Baseline=0 • pH=0 • Fouling=0 • Ionic=0
4b
Stage 4b — AWGN + Baseline
Cubic+sigmoid+4-physical baseline • Capacitive spike, drift, step, oscillation • 5K samples
▸ Noise Profile (stage "4b")
\[I_\text{bg}=aE^3+bE^2+cE+d+\text{sigmoid}\]
+ Probabilistic add-ons (25%/40%/20%/30%):
Capacitive spike • Exponential drift
Step discontinuity • Low-freq oscillation
AWGN=1.0 • Baseline=1.0 • pH=0 • Fouling=0 • Ionic=0
4c
Stage 4c — + pH Shift
Analyte-specific E₀ shift • no wrap-around • 5K samples
▸ Noise Profile (stage "4c")
\[\Delta E_{0,\text{HQ}}=s_\text{HQ}\cdot(\text{pH}-\text{pH}_\text{ref})\] \[s_\text{HQ}\sim\mathcal{U}(-60,-40)\;\text{mV/pH}\]
Separate slope for DA • E₀ shifted directly, no roll()
AWGN=1.0 • Baseline=0.8 • pH=1.0 • Fouling=0 • Ionic=0
4d
Stage 4d — + Ionic Effect
HQ amplitude suppression + small E₀ shift • Blood serum etc. • 5K samples
▸ Noise Profile (stage "4d")
\[A_{\text{HQ}} \leftarrow A_{\text{HQ}}(1-r),\quad r\sim\mathcal{U}(r_{\min},r_{\max})\] \[E_{0,\text{HQ}} \leftarrow E_{0,\text{HQ}} + \mathcal{U}(0,\,\Delta E_{\max})\]
HQ-specific only — DA unaffected
AWGN=1.0 • Baseline=0.8 • pH=0.8 • Fouling=0 • Ionic=1.0
4e
Stage 4e — + Fouling
Amplitude attenuation + positive E₀ shift • Electrode fouling model • 5K samples
▸ Noise Profile (stage "4e")
\[A \leftarrow A\cdot e^{-k_{\text{eff}}\cdot t_n},\quad k\sim\mathcal{U}(k_{\min},k_{\max})\] \[E_0 \leftarrow E_0 + \Delta E,\quad\Delta E\sim\mathcal{U}(0.022,0.050)\cdot t_n\]
Separate k_eff for HQ and DA • Unidirectional E₀ shift
AWGN=1.0 • Baseline=0.8 • pH=0.8 • Fouling=1.0 • Ionic=0.8
4f
Stage 4f — Full Noise
All noise sources at maximum • Real laboratory conditions • 12K samples
▸ Noise Profile (stage "4f")
All noise sources p=1.0:
AWGN • Baseline (cubic+sigmoid+4 add-ons)
pH shift • Ionic effect • Fouling
Capacitive scan spike (25%)
AWGN=1.0 • Baseline=1.0 • pH=1.0 • Fouling=1.0 • Ionic=1.0