Accuracy, Uncertainty and Explainability of Mixed Input Neural Network Models for Prediction of Missile Aerodynamic Loads

Prediction models for aerodynamic loads of missile configuration were developed using CNN (Convolutional Neural Network)-based Mixed Input Neural Network (MINN), inputs of which are image in the first input layer and flow conditions in the second input layer. The signed distance function was used to convert missile geometry to image data with various resolutions. Aerodynamic dataset was generated by using Missile DATCOM. To check the uncertainty due to the randomness of trained deep neural network models, box plots of prediction errors from independently trained multiple models were generated and compared. Grad-CAM was used to provide explainability to missile configuration images. It was confirmed that predictions of the neural network model become robust as the size of dataset increases. MINN1, which only uses image resolution of whole missile geometry was found to be unable to accurately predict effects of the nose type on axial force. Therefore, the missile nose type information was explicitly provided to the second input layer in MINN2. Then, a CNN classification model was developed for nose type prediction using signed distance function image around the missile nose. The accuracy of the nose type classification CNN model was 100% for 64² or higher input image resolutions. CNN + MINN2 architecture, which provides MINN2 with the predicted nose type by the CNN model, was built and compared with MINN1 and MINN2. MLP (Multi-Layer Perceptron) was also used for comparison with the MINN models as a reference. Both MINN2 and CNN + MINN2 are confirmed to be very accurate having relative errors less than 5%.