Predicting Velocimetry Curves from Photonic Doppler Velocimetry (PDV) Signals using Neural Networks

Photonic Doppler Velocimetry (PDV) is a popular diagnostic technique for generating particle (or free surface) velocity profiles during shock compression experiments. However, the signal gathered by the probe requires post-processing using software that involves either human intervention for different settings (SIRHEN) or brute force averaging (HiFiPDV) to produce the final velocity profile. For experiments with multiple PDV probes, it becomes difficult to perform a consistent and rigorous analysis to identify and isolate the effects of material microstructure, especially for heterogeneous materials.

The goal of this work is to establish velocimetry curve consistency between multiple PDV probes through use of a Neural Network (NN) model, thereby reducing the uncertainties and errors associated with the current processing methods. The raw PDV signal is pre-processed and fed into the NN model with hidden layers, which outputs the predicted velocimetry curve. The work specifically utilizes LSTM (Long Short-Term Memory) Recurrent Neural Networks (RNNs). The method is expected to predict optimum outputs for different materials if it is fed a wide range of training data, because it learns from past inferences. The RNN model description and results obtained to date will be presented.