Using Recurrent Neural Networks to Reconstruct Temperatures from Simulated Fluorescent Data for use in...

Area of Science:

• Biophysics
• Microfluidics
• Machine Learning

Background:

• Biological systems often require precise temperature control for accurate study.
• Existing temperature sensors struggle to provide both high accuracy and spatial resolution in microfluidic devices.
• Microfluidics are essential for isolating and studying biological systems.

Purpose of the Study:

• To introduce a novel neural network for high-accuracy, pixel-level temperature mapping in fluorescent images.
• To overcome limitations of standard fitting practices and existing machine learning methods for temperature determination.
• To enable enhanced temperature resolution in microfluidic studies of biological systems.

Main Methods:

• Development of the Multi-Directional Fluorescent Temperature Long Short-Term Memory Network (MFTLSTM).
• Utilizing heat diffusion properties within fluorescent images for temperature calculation.
• Validation with both simulated and experimental data from 3D printed microfluidic devices.

Main Results:

• MFTLSTM achieved an accuracy of ±0.0199 K RMSE with simulated data (298 K–308 K).
• Experimental data (290 K–380 K) showed an accuracy of ±0.0684 K RMSE.
• The network accurately calculates temperature at every pixel in fluorescent images.

Conclusions:

• The MFTLSTM network significantly improves temperature measurement accuracy and spatial resolution in microfluidics.
• This technology has the potential to advance research in biological systems requiring precise thermal control.
• The method offers a new standard for temperature analysis in microfluidic applications.