Study on neural network setting proper samples in dynamic light scattering particle size analysis

Dynamic light scattering is often used to detect small particles, such as in industry and medicine. A typical ill-posed problem needs to be solved to recover PSD by inverting ACF data , which is a difficult problem of DLS. when DLS is used for detecting small particles in transformer oil, it is difficult to accurately recover PSD using traditional algorithms. Generalized regression neural network(GRNN) has been proved to be applicable to solving ill conditioned equations in Dynamic light scattering method. However, accurate inversion relies on proper training sets closely matching measured particle size to avoid large errors. Generating numerous samples for multimodal distributions is time-consuming. This study investigates how sample setting range affects GRNN inversion accuracy during training.The experimental system was self-built, using 362.2nm and 806.9nm polystyrene mixed diluted lotion as selected samples. The training sets were centered around theoretical particle sizes, with range variations of 10nm, 30nm, 50nm, and 100nm. The GRNN was trained using these sets, and the experiment's light intensity autocorrelation data was input into the neural networks to obtain particle size distributions and bimodal peak particle sizes. All the sample set were achieved by measuring ACF of 362.2nm and 806.9nm polystyrene suspension on a self-built DLS experiment system. These findings indicate that closer proximity between the sample range used in neural network training and the actual situation leads to more accurate inversion results, demonstrating the network's ability to accurately invert bimodal samples. Furthermore, the accuracy improves with more realistic training set settings. In practical measurements, combining regularization methods with this approach can enhance particle size analysis accuracy.