JavaScript isn't enabled in your browser, so this file can't be opened. Enable and reload.

1 of 22

Fluorescent �multimodal nanosensor�of heavy metal ions �based on carbon dots

Olga E. Sarmanova, Kirill A. Laptinskiy,

Galina N. Chugreeva, Sergey A. Burikov, Tatiana A. Dolenko

Lomonosov Moscow State University

2 of 22

Pb Hg Cd Cu Co Fe Cr Ni Zn

3 of 22

Method	Simultaneous detection of several elements	Non-destructive or doesn’t need sample preparation	Cost	Detection limit
Atomic Absorption Spectrometry	✔	–	$$	~ 10^-5 g/L
Radiochemical methods	✔	–	$$$	~10^-10 g/L
Electroanalytical methods	–	–	$$	~ 10^-6 g/L
Optical spectroscopy	✔	✔	$	~ 10^-8 g/L

^{*S. Morais, F. G. e Costa and M. de Lourdes Pereira. Heavy Metals and Human Health. Environmental Health – Emerging Issues and Practice}

Current state

4 of 22

Carbon dots (CD) -based nanosensors

Size, nm	Fraction content in suspension, %	Zeta-potential, mV
30±1	99.8	-39± 0.9
398±2	0.2	-39± 0.9

Image of CD in a scanning electron microscope.

Intense and stable fluorescence (FL)
Sensitivity of the CD FL to changes in the environmental characteristics
Non-toxicity and biocompatibility

[*] Molaei, M. J. (2019). Carbon quantum dots and their biomedical and therapeutic applications: a review. RSC Advances, 9(12), 6460–6481.

[*]

𝜆_ex = 365 nm

5 of 22

Our aim

To develop an approach to create

a multimodal nanosensor of metal ions, based on neural networks application to the fluorescence spectra

of carbon dots aqueous suspensions.

6 of 22

Carbon dots (CD) -based nanosensors

Size, nm	Fraction content in suspension, %	Zeta-potential, mV
30±1	99.8	-39± 0.9
398±2	0.2	-39± 0.9

Image of CD in a scanning electron microscope.

Intense and stable fluorescence (FL)
Sensitivity of the CD FL to changes in the environmental characteristics
Non-toxicity and biocompatibility

[*] Molaei, M. J. (2019). Carbon quantum dots and their biomedical and therapeutic applications: a review. RSC Advances, 9(12), 6460–6481.

[*]

𝜆_ex = 365 nm

7 of 22

Sensitivity to Cations

8 of 22

Why �Machine Learning?

Unclear fluorescence mechanisms
Different synthesis methods ->

different structures in the CD composition

No analytical models

describing CD-environment interaction

Nonlinear dependence of FL

on measured parameters

9 of 22

Fluorescence Spectroscopy

Fluorescence emission spectrum of CD under excitation at different wavelengths

Excitation-emission matrix

of CD's suspension fluorescence

Columns

Rows

Excitation

spectra

Fluorescence

spectra

10 of 22

Dataset - I

Sample: 41 FL spectra of CD suspension, corresponding to 41 excitation wavelengths in 250 – 450 nm spectral range with 5 nm increment.

Spectrum: 500 spectral channel (features) in 250-750 nm spectral range with 1 nm increment.

Outputs:

Cu²⁺, Ni²⁺, Cr³⁺, NO₃^- concentrations, mM

Range: 0 to 4.95 mM,

0.05 mM increment

1000 samples

11 of 22

Dataset - II

Sets:

Training set (70%), validation set (20%), test set (10%)

Splitting way:

Random split due to the uniform concentrations grid.

Stopping criterion:

If MSE on the validation set did not decrease during 100 epochs, stop

12 of 22

Sensitivity to Cations

Channel-by-channel statistics of the fluorescence spectra array for CD aqueous suspensions in the presence of a single salt. The red line is the average value for the array in each spectral channel (feature), the black dotted line is the standard deviation of the feature value.

13 of 22

PCA for Preliminary analysis

Score plots samples projections into the space of the first three principal components.

Cu²⁺ and Cr³⁺ cations: clear clusters, formed by the samples with similar concentrations of the corresponding ions.

Ni²⁺ cation: impossible to highlight clusters, as fluorescence spectra of suspensions with similar or equal Ni²⁺ concentrations differ greatly.

Conclusion: the error in determining the concentration of Ni²⁺ using CD will be maximal.

�

14 of 22

Approaches to Solve Inverse Problem- 1

Multilayer Perceptrons (MLP)

Approach	Wavelength of FL Excitation	Sample size
MLP_1W	350 nm	[1x500]
MLP_3W	250, 350, 450 nm	[1x1500]
MLP_41W	41 excitation wavelengths	[1x20500]

15 of 22

Approaches to Solve Inverse Problem - 2

Convolutional neural networks (CNN)

Approach	Number of channels	Sample size
CNN_1D	41	[41x1x500]
CNN_2D	1	[1x41x500]

Sample:

16 of 22

Basic models

+ 5-fold cross-validation

+ 5 initializations of networks

25 networks per case

17 of 22

Results Comparison. Basic models.

Choose 2D_CNN

18 of 22

Feature maps

19 of 22

Data Augmentation

Wavelength drifting

x_noise = x_init + noise_level * rand_uniform

x_noise = x_init + noise_level * sqrt(x_init) * rand_poisson(λ = x_init)

20 of 22

Results Comparison. Augmentation

21 of 22

Conclusion-I

Different sensitivity rates of CD fluorescence toward cation types dramatically influence the results of NNs application;

2D CNNs enabled achieving a minimum RMSE -> as they ‘consider’ additional information about the mutual arrangement of spectral channels, both in the fluorescence and excitation spectra;

22 of 22

Conclusion-II

We conducted research on the creation of a fluorescent carbon nanosensor for the simultaneous determination of heavy metal ions concentrations using machine learning methods.
Fluorescence of CD prepared via hydrothermal synthetic route is extremely sensitive to changes in ions concentrations.
The developed method enables simultaneous determination of the concentrations of heavy metal ions Cu²⁺, Ni²⁺, Cr³⁺ with an RMSE of 0.28±0.03 mM, 0.79±0.04 mM, 0.24±0.02 mM, respectively.