Use of Machine Learning with Temporal Photoluminescence Signals from CdTe Quantum Dots for Temperature Measurement in Microfluidic Devices

Abstract

Because of the vital role of temperature in many biological processes studied in microfluidic devices, there is a need to develop improved temperature sensors and data analysis algorithms. The photoluminescence (PL) of nanocrystals (quantum dots) has been successfully used in microfluidic temperature devices, but the accuracy of the reconstructed temperature has been limited to about 1 K over a temperature range of tens of degrees. A machine learning algorithm consisting of a fully-connected network of seven layers with decreasing numbers of nodes was developed and applied to a combination of normalized spectral and time-resolved PL data of CdTe quantum dot emission in a microfluidic device. The data used by the algorithm was collected over two temperature ranges: 10 K to 300 K, and 298 K to 319 K. The accuracy of each neural network was assessed via mean absolute error of a holdout set of data. For the low temperature regime, the accuracy was 7.7 K, or 0.4 K when the holdout set is restricted to temperatures above 100 K. For the high temperature regime, the accuracy was 0.1 K. This method provides demonstrates a potential machine learning approach to accurately sense temperature in microfluidic (and potentially nanofluidic) devices when the data analysis is based on normalized PL data when it is stable over time.

Keywords: Photoluminescence; fluorescent lifetimes; machine learning; quantum dots; thermometry.

Figure 1.

Schematic of COOH functionalized CdTe nanoparticle inside open source PR48 resin. QDs were purchased as a powder from PlasmaChem and mixed into PR48 resin as described in the body of the text. Some of the key compounds in PR48 resin include: Genomer 1122, Ebecryl 8210, and Sartomer SR 494.

Figure 2.

Optical experimental setup (a). The laser diode excites photoluminescence (PL) from the sample, which is placed in a cryostat or on a heating stage. The PL is collected and focused onto a spectrometer. For spectral scans the spectrometer analyzes the PL for wavelength dependence using a photomultiplier tube (PMT) detector or CCD detector (not shown). For time-resolved PL, the laser is controlled by a pulse generator which also sends the start signal to the TCSPC module, the spectrometer is set to the peak wavelength, and the PMT signal triggers the TCSPC module to stop. Microfluidic samples (b). CAD model sent to printer and PR48-resin printed device with CdTe impregnated PR48 resin filling channels fluorescing under 532 nm light. Empty channels for the thermocouple (TC) and heating (Galinstan) are shown.

Figure 3.

Representative normalized data from 10 K to 300 K. (a) Spectral data (wavelength-resolved photoluminescence). (b) Time-resolved photoluminescence obtained via time-correlated single photon counting (TCSPC).

Figure 4.

Peak intensity of CdTe dots embedded in a polymer matrix, showing stability of the PL signal over three days (D1-D3), and multiple tests on the same day (T1-T3).

Figure 5.

Complete set of training data with interpolations for the low temperature regime. (a) Spectral data. (b) TCSPC data. Aside from the logarithm preprocessing step for the TCSPC data mentioned in the text, the plots shown in Fig. 2 are horizontal slices of these images at 10, 100, 200, and 300 K.

Figure 6.

Complete set of training data with interpolations for the high temperature regime. (a) Spectral data. (b) TCSPC data.

Figure 7.

Plot of loss functions of training and testing sets during neural network training for the low temperature regime.

Figure 8.

Performance of the neural networks on training, testing/validation, and holdout sets for (a) the low temperature regime and (b) the high temperature regime.

Figure 9.

Performance of neural network predictions with noise artificially added to the optical data for (a) the low temperature regime and (b) the high temperature regime.