چكيده لاتين
Dynamic Light Scattering (DLS) is a non-contact and non-destructive optical technique used to measure the size of suspended particles in colloidal environments. It is based on analyzing the fluctuations in light intensity caused by the Brownian motion of particles in a fluid. DLS has widespread applications in the pharmaceutical, food, environmental, paint, and nanotechnology industries. However, analyzing DLS data becomes challenging when particle sizes are large or the fluid concentration is high. The aim of this study is to enhance the accuracy and speed of particle size estimation by applying deep learning to directly analyze raw DLS signals. To this end, a laboratory system was designed and implemented, consisting of a 532 nm second harmonic Nd:YAG laser source, mechanical chopper, detector, digital oscilloscope, and cuvette. Microsilica particles were used as the sample material, classified into nine defined size ranges using a standard sieve shaker, and then suspended in solution. Experiments were conducted by varying parameters such as laser power, sample-to-detector distance, detector shielding to reduce environmental noise, oscilloscope sampling rate, and cuvette geometry. The output light intensity signals were recorded as time series and directly used in the analysis process. In the processing phase, two deep neural network models—classification and regression—were designed and trained. The input signals, in their entirely raw form and without any preprocessing, were fed into the networks. For training, a database of over 〖9×10〗^6 numerical light intensity signals was used. Results showed that the classification model achieved approximately 96% accuracy, while the regression model attained a mean absolute error of approximately 0.32 micrometers, indicating low error in estimating the hydrodynamic diameter of particles. The use of neural networks significantly reduced the signal analysis time. Furthermore, both models demonstrated acceptable performance when applied to signals from polydisperse samples. Overall, this study shows that employing neural networks with raw data not only eliminates the need for complex numerical analyses but also improves accuracy, reduces computational time, and enhances the stability of data analysis. This approach represents a step toward automating DLS data interpretation and developing precise laboratory tools capable of handling complex datasets.
