Machine Learning Approaches for Inverse Problems and Optimal Design in Electromagnetism
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThe work is complete and well-exposed. The machine learning approaches are detailed yet concise, and the techniques are well-described. The results are clearly articulated.
Author Response
The work is complete and well-exposed. The machine learning approaches are detailed yet concise, and the techniques are well-described. The results are clearly articulated.
We thank the reviewer for the appreciation of our work.
Reviewer 2 Report
Comments and Suggestions for Authors1. This article explores some basic deep-learning models for solving electromagnetic problems. However, it lacks innovation in deep learning theory or addressing electromagnetic problems. Please improve it.
2. This work compared the use of shallow neural networks with convolutional neural networks and support vector machines for the same electromagnetic problem but lacks technical contribution.
3. The author initializes the input data of the neural network with a special treatment for the characteristics of the electromagnetic problem. The scope of the application is not good enough.
Author Response
Please look at attached file
Author Response File: 
Author Response.pdf
Reviewer 3 Report
Comments and Suggestions for AuthorsDear Authors,
The paper deals with the resolution of an electromagnetic indirect problem via machine learning algorithms. Given the computational effort that is usually required to solve the problem, the methods proposed offer very interesting approaches to overcome long computational time. The results show general clear agreement between the predicted and true values. For immediate comprehension of the Figs. 7, 8, 9, I strongly recommend the Authors the add the colour legend inside the plots to visualize the true and predicted values.
Concerning the general discussion, I would suggest to the Authors to address the following points:
1) Computation time: it would be interesting to have an estimation of the computation time required to solve the indirect problem via the ML approaches selected by the Authors to assess the time performance of each method.
2) Training of the NN: it is stated that the training is done either with data produced with FEM or analytical calculations. Usually, NN needs huge numbers of training data to become accurate in finding the solutions of the indirect problem. In the manuscript the different models have been assessed with 26000 examples generated by FEM and then divided according to the 70/30 general rule. Could the Authors explain more about the generation of 2600 FEM data? Can this data also be used to solve problem with different geometry or there is the need to generate FEM data for each electromagnetic problem to be solved? It would interesting to address this point in the paper.
Thank you very much.
Author Response
Please look into the attached file
Author Response File: Author Response.pdf
Round 2
Reviewer 2 Report
Comments and Suggestions for Authors1. This article employs an adaptive learning rate algorithm for a specific electromagnetic problem. Improved accuracy and speed compared to other algorithms, It is innovative.
2. This article shows a lot of explorations and explanations for the solution of the electromagnetic problem. The final proposed model is adaptable for similar problems which is meaningful.
