Neural-Network-Based Nonlinear Model Predictive Control of Multiscale Crystallization Process
Round 1
Reviewer 1 Report
The paper “Neural Network-based Nonlinear Model Predictive Control of Multiscale Crystallization Process” describes a methodology to use deep learning and image analysis in a feedforward manner to control the mean crystal size. Authors have tested the effectiveness of their methodology using alum cooling crystallization. Results are promising but there is a lack of scientific discussion in the paper. This must be improved.
1. Introduction
Overall, it contains relevant information, but the readability is very poor. Sentences are disconnected from each other which makes it hard to follow. There is no clear statement of the state-of-the-art, the challenges faced in the field, or the problems that this work is trying to tackle. It would be beneficial for the reader’s understanding if the text is revised, and the authors relate the previous works that are mentioned with their own work in a clear way.
Example: On lines 32-35, there is a cause-consequence relation between two statements (Statement 1 - Hence - Statement 2), with no clarity whether it highlights a positive or a negative aspect. The model cannot be implemented into the controller (seems like a negative) and because of that its order is reduced and that lowers computational costs (seems like a positive). It states it captures the dominant dynamic: is this good or bad? Is it enough?
Also, I believe that “dominate dynamic” should be “dominant dynamic”.
Many abbreviations are not defined. Even if they seem obvious, it is good practice to define them the first time they appear in the text.
Line 54: Authors mention “fine scale information”. What exactly is meant by fine scale?
2. Multiscale dynamical model description
The dynamic evolution of a crystallization process that is presented by the authors in a macroscopic model does not mention any nucleation variable in the equation. What are the considerations to assume that? Is this a seeded crystallization or is it supposed to monitor a continuous crystallization process? If seeded, how do the seeds size distribution affect the number density function?
Microscopic model is referred to a solid-on-solid model reference. Even though this model was not developed in this work, it is good practice to provide a brief (1-2 sentences) description of the physical interpretation and assumptions from the model.
Line 100: What is the physical meaning of this so-called “migration rate”?
Figure 1 is not explained in the text, but simply referred to. The explanation that comes after this figure is confusing and needs to be rewritten.
3. Model and controller design of NMPC
Line 126: Authors mention “detection and extraction”. What do they mean by extraction?
How was the training set performed?
4. Experimental results
Line 185-189: Details about the crystallization are provided. They should be in a Material and Methods section, not on the results. Furthermore, the details are not enough to understand the process. Is alum added, dissolved and then recrystallized? Or are the commercial crystals only partially dissolved and then have their size controlled through the model? What are the temperatures used?
Figure 5 shows predicted mean size vs measured mean size. Authors should quantify goodness of fit, not just show the plot.
The 2nd paragraph of the results section presents all the figures and shortly describes whatever is displayed in the figures. However, results are not discussed, just presented. There is only one paragraph more after the figures and it is mere 4 lines that provide some considerations, not discussion.
Even though the results seem good, quantification of the fittings are missing. Experimental data points do not show any error bars, standard deviation, or confidence intervals that really attests the accuracy of the measurements.
As a strategy to control crystal size, is controlling mean size enough for a crystallization process? How is this mean size compared within the whole crystal size distribution? Have the authors measured the crystal size distribution?
Please discuss crystal mean size, particle size distribution and how does that affect the feasibility of the control.
How is the control influenced by particle shape? Can it be generalized, i.e., used to other systems under crystallization, if crystals are needle-like, for example?
Figures 7 and 9, and 6 and 8 could be coupled either side-by-side or in a single plot, to allow for easier comparison between PID and NMPC.
5. Conclusions
Since the discussion is poor, conclusions are not well-supported.
Author Response
We would like to thank the reviewers for giving us constructive suggestions which would help us both in English and in depth to improve the quality of the paper. The attached file is a point-to-point response to the two reviewers’ comments. In addition, we mark all the changes in the revised manuscript.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The paper is well-written and structured. The introduction and related works sections provide sufficient background for the reader. The results are presented clearly, and the paper is easy to read and understand. The paper focuses on developing a control strategy for multiscale batch cooling crystallization. The image analysis and statistical methods were adopted to create a nonlinear model predictive control method to obtain the mean size of the crystal population.
The authors claim first to present the “deep learning-based on-line measurement method of micro characteristic”. However, there is almost nothing in the paper about the mentioned method. Moreover, the reader may suggest that the deep-learning component is used from [25] (see lines 125-128). If the authors indeed have developed the “deep learning-based on-line measurement method,” I expect to get its description in the paper, including the training, validation, cross-validation etc. Moreover, deep-learning methods usually require large training datasets, so I’d expect a detailed description of such datasets as well.
Hence, if the authors conclude that they have developed a method for growth-process control, I would be satisfied; their experiments can prove the validity of such a method, and the paper would lie in the material science/chemistry field; though they claim to develop the “deep learning measurement method” I’d like to read more about that, and this is more or less in the computer science area. Hence, I suggest adding more about the seep learning method or shrinking the paper to the mentioned field.
Maybe the article is good from the chemical or material science point of view, but I cannot correctly evaluate it from the computer science point of view:
· The dataset is not described: the size, division on a training set and validation set, etc.
· The data preprocessing, if any, is not described.
· The activation function and the ANN architecture are not substantiated (why exactly are these functions, 4 inputs, one hidden layer with 10 neurons, etc.).
· The validation (and cross-validation!) is not mentioned.
· The accuracy metrics and performance evaluation are not presented.
Hence, the reproducibility of the present research from the computer science point of view is in doubt.
In line 218, the authors claim that “The developed NMPC has been tested and proven to perform well with significant improvements over PID controllers …”. However, the paper does not show the comparison of the proposed NMPC and the traditional PID controllers. So, this conclusion is not supported by the results presented.
Besides, from Fig. 5, one can suggest that a traditional regression technique or polynomial approximation could also be used for predicting the mean crystal size. It is unclear from the paper whether such methods have been used in other studies; if yes, what are their disadvantages etc. In a few words: why do the authors use feed-forward NN instead of a simpler method?
Minor spell-checking is needed, e.g.:
1. Line 26: missing subject - A more systematic (what?) to the model and control of multiscale process is needed [4].
2. Line 102: The computation of the growth rate cannot be computed
Author Response
We would like to thank the reviewers for giving us constructive suggestions which would help us both in English and in depth to improve the quality of the paper. The attached file is a point-to-point response to the two reviewers’ comments. In addition, we mark all the changes in the revised manuscript.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Introduction and methodologies (model and experiments) were greatly improved.
Description of Figure 1, however, is still very confusing. Some numbers were added, which are not straightforward.
Discussion of the results could still be improved. For instance, parts of the answer provided by the authors to point (12), on controlling size vs size distribution could be a point of improvement to the text.
In the new text added, some minor verbal agreement mistakes are present and need to be revised.
Author Response
Please see the attached file.
Author Response File: 
Author Response.doc
Reviewer 2 Report
The revised version of the paper has been improved. However, it is still out of computer science scope and hence a bit out of my area of expertise.
In this case, I suggest making an accent on control of the crystallization process rather than on neural networks. The aim of the paper, viz. "to provide an integrated control strategy for multiscale batch cooling crystallization process", supports my suggestion. Besides, the authors answer that they "do not improve the feedforward neural networks and image analysis method using deep learning methods". Thus, they have used the ANN models (with a pre-trained deep learning one, as I can understand) just as a tool to build a control strategy for the crystallization process.
I suggest removing Fig. 10 and leaving only the explanation in the text (lines 246-250). In this case, Fig. 10 is confusing: there is no visible difference between PID and NMPC in Fig. 10. Indeed, it's hard to distinguish 5.3 μm and 9.2 μm in the figure with the scale from -180 to 20. Moreover, what does it mean Error/μm in this figure?
Author Response
Please see the attached file.
Author Response File: Author Response.doc
