| Name: | Ethan Regal |
|---|---|
| Email: | regal002@gannon.edu |
| Office: | 446 |
| Home Institution: | Gannon University |
| Project: | Deep Learning for Quality Prediction in Metal Additive Manufacturing |
I began reading relevant papers to the topic, spent a great deal of time narrowing my interests and aligning my goals and objectives for the summer. Have decided to work solely in predicting emission values in PBF. Will supplement this approach with many different models and an explainable AI technique. Made the presentation as well.
This week I made four different models. A regression neural network with features of laser power, laser speed, and layer number to predict the average emission value of a layer. I then proceeded to design time series models as there is an expectation for time series analysis to perform more effectively. I designed one ARIMA model, a regression neural network redesigned with time series components, and an LSTM model. There seems to be a decrease in loss on the testing set as complexity increases. I wish to design one more model before beginning the explainable AI approach. Among this work, I have also been closely reading other works with similar pursuits.
This week I had developed one more model that uses features of laser power, laser speed, and five sequential layer emission values and predicts the next layer value. This model was trained using all training data, and performs reasonably well. Furthermore, I had begun to implement explainable AI techniques. I had developed a partial dependence plot for the model with indpendent features and ALE plots for the remaining models with correlated features. More work will go into interpreting the ALE plots as well as understanding and implementing more XAI techniques.
This week I had done an immense amount of literature review surrounding the theory behind explainable AI techniques. I had thoroughly read through the theory surrounding the SHAP and LIME techniques. I plan on comparing both of these techniques and determining their capabilities.
I had spent this week implementing the LIME and SHAP techniques to each of the five models. I had also developed global explanations from the results of each. Along with this, I had also done research into implementing the technique of Integrated Gradients and thoroughly looked into the theory regarding the technique. After I had wrote a function to determine an explanation of a prediction with the use of integrated gradients. Given the similarity of the results to the other techniques, there is some reason to believe I may have written the function correctly.
SHAP, LIME, and IG are all performing within expectations. That being said, the explanations still differ in sometimes great capacity for each, so I began researching and successfully implemented metrics that determine how "quality" an XAI technique is. To form richer results, I had added one more model in even greater complexity - a hybrid regression neural network. Such a model computes a pediction from time sensitive features and a prediction from static features. Another neural network then "combines" these predictions to deliver a final prediction. This model performs better than the previous, at a testing loss of ~0.85 RMSE. Due to the complexity, applying the XAI techniques proved quite challenging, but was ultimately completed.
This week was primarily spent introducing the physics based loss functions into the model. Two functions were formed based on observations from the training data, one function that penalizes the model for exceeding the maximum emissivity of a given laser power and laser speed and the other which penalizes the model for failing to exceed a moving average of the five previous layers. Whether these functions were successful or not was determined using the ALE function. Current results indiate that these two physics based loss functions do in fact enforce physics behavior of the model.
This week was spent preparing the presentation and compiling all the results. There was a significant challenge in selecting which results and background to cut in order to remain in time constraints. At the end of the week I had presented my project and while I was unable to present everything I had wanted, I believe I conveyed the primary findings of the project.
All effort this week consisted of writing my technical report. I met with my mentor on Monday to discuss the structure, and promptly began work afterwards. To present all the results in a coherent matter within a few days proved to be challenging but was utimately done. I am proud of my report and am extremely happy with my time spent with DIMACS this summer. Future work will be done with my mentor in writing a more technical paper intended for publication. I thank DIMACS and Rutgers University for this opportunity and thank the NSF for graciously funding this research via the NSF grant CNS-2150186. Finally, I give special thanks to my mentors for providing me extremely valuable insights and comments during the course of this project.