Porosity Prediction of Granular Materials through Discrete Element Method and Back Propagation Neural Network Algorithm
Round 1
Reviewer 1 Report
This manuscript presents a porosity prediction model of granular materials using DEM and BP neural network algorithm. The DEM simulation results serve as training and testing data for the BP neural network model. Experiments on steel balls are used as validation for the proposed model. I have several major concerns that must be resolved before this manuscript becomes publishable on Applied Science. The detailed comments can be found as follows.
Some grammar issues. I could never find all of them. I would suggest the authors to double check on their own. P1, L12: “compute the porosity of a granular material which contain a wide range particle sizes or shapes” should be “compute the porosity of a granular material which contains a wide range of particle sizes or shapes”; P1, L39:”A lot of work have done to study” should be “A lot of work has been done to study”; P3, L9:”Machine learning algorithms requires large sample size to” should be “Machine learning algorithms require large sample size”; P7, L24:”is to processing the original input data” should be “is to process the original input data”. BP Neural Network algorithm. What does BP stand for? Please specify. P3, L12:”as shown in Error! Reference source not found”. Please correct. What is the influence of static pressure applied for the compaction (fixed as 600 KPa)? My first major concern. There are analytical (or semi-analytical) models to predict the packing density. Considering the spherical particle assumption in the paper (which is really not complicated), what are the advantages of DEM and NN? P2, L 26:”the effect of aggregate on the porosity of asphalt mixture is significantly greater than that of asphalt.” Please explain this in the text. Table 1 and 2. What AC, AM, OGFC, SMA stand for? Why are two gradation types, gradation-13 and gradation-16, selected in the paper? How about other gradation types? My another major concern. Please specify the necessary parameters of the DEM model. For instance, how long does one DEM simulation take? How many particles are in the DEM? What are the boundary conditions? These properties are very important to analyze the simulation results. The procedure for standardizing the input and output variables is not clear. For instance, what is y in the Eq. (3)? The validation of NNM and DEM against experiments is interesting. What is the size of the container? Can you give a specific particle size distribution of the steel balls? My last major concern. P14, L29:” Particle shape is a minor factor compared with gradation and its impact can be considered by developing corresponding correction coefficient”. That is a huge conclusion. I cannot agree with it. Particle shape, as far as I know, is also a very important factor of the sample porosity.Author Response
Please see the attachment.
Author Response File: 
Author Response.docx
Reviewer 2 Report
In this paper, the authors proposed a method for predicting the porosity of the granular media based on a machine learning algorithm, called BP Neural Network Algorithm. The algorithm training was done using the Discrete Element Method (DEM) for producing the adequate training data. The method was tested by the numerical and experimental models. Although I don’t know well the machine learning, I found the paper is very interesting and I recommend a publication in Applied Sciences after having clarified the following issues:
- The DEM training data is based on the Chinese Technical Specification for Construction of Highway Asphalt Pavements JTG F40-2004, and Table 1 and 2 show it, but some parameters like AC, AM… were not defined and are not cleared for the readers. Please define them.
- The DEM samples were compressed by applying a pressure of 600 KPa. Why this value?
- In this study, the size ratio of particles doesn't exceed the value of 2. However, in the reality, the size polydispersity can be very large. What is the effect of large polydispersity?
- The shape of particles can also influence the porosity (specially elongated particles). Since the authors considered only the particle volume fraction as the input data, how should we introduce the shape effects?
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Reviewer 3 Report
BP was used as an abbreviation throughout the manuscript. I assume BP means Backpropagation which is only the way by which the neural network is trained. The kind of neural network used was however not highlighted (linear, convolutional etc). Critical information such as the learning rate, the batch size is missing. No justification is giving for the activation functions used.
Critical information is missing.
Author Response
Please see the attachment.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The paper sounds good.
