Data-Driven Prediction of Crystal Size Metrics Using LSTM Networks and In Situ Microscopy in Seeded Cooling Crystallization

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The article “Data-Driven Prediction of Crystal Size Metrics Using LSTM Networks and In Situ Microscopy in Seeded Cooling Crystallization” is devoted to a data-driven modeling framework for predicting image derived crystal size metrics in seeded cooling crystallization using Long Short-Term  Memory (LSTM) neural networks. The article is written in good language, contains a significant amount of experimental material and is beautifully illustrated. A large number of works by other authors are cited, with appropriate references. At the same time, it should be noted that studies  “in situ” crystallization processes were also carried out using other observation methods, in particular AFM, for example, in [Piskunova, N. N. Non-reversibility of crystal growth and Dissolution: Nanoscale direct observations and kinetics of transition through the saturation point. Journal of Crystal Growth. - 2024. -V. 631. -127614]. Perhaps such studies should also be noted in the literature review.

In my opinion, this model is promising, but in its current form it can only be recommended for use in very simple systems with a minimum of variables. For example, the authors take into account the introduction of seeds and calculate supersaturation, but do not take into account the crystallographic orientation of the seeds, which is relevant, for example, during industrial growth of large KDP/ADP crystals from aqueous solutions using the high-speed crystallization method. An even greater number of variables appear when growing single crystals from high-temperature melts using the Czochralski or Naken-Kyropoulos methods, where in addition to supersaturation and seeding, variables appear in the form of rotation rates and crystal movement. The authors should dwell in more detail on the prospects for using this technology in industrial crystal growth.

In general, this article can be useful for researchers involved in predicting the crystallization of new materials under simple conditions. With these minor revisions, the article can be recommended for publication in the “Processes’.

Author Response

Reviewer Comment 1:

The reviewer suggests noting additional studies using alternative “in situ” crystallization observation methods, specifically mentioning atomic force microscopy (AFM) studies such as Piskunova (2024).

 

Authors’ Response:

We thank the reviewer for highlighting the significance of other "in situ" methodologies employed in crystallization studies. Indeed, atomic force microscopy (AFM) offers valuable insights into crystal growth and dissolution processes at the nanoscale, complementing our chosen in situ microscopy technique. As suggested, we will enhance our literature review by including relevant AFM studies, specifically referencing the work by Piskunova (2024), to provide readers a broader context of methodologies used to monitor crystallization processes. The work by Piskunova is cited under number [45] and text is added at rows 97-99 “While this work applies LSTM networks to predict macroscale crystallization behavior using in situ microscopy data, similar data-driven approaches could also support modeling of kinetic phenomena at the nanoscale.

Reviewer Comment 2:

The reviewer highlights that, while promising, the presented model in its current form appears suitable mainly for simple crystallization systems and does not account for important industrial-scale crystallization variables such as crystallographic orientation, rotation rates, and crystal movement, particularly relevant in KDP/ADP crystal growth and high-temperature crystal growth methods.

 

Authors’ Response:

We fully acknowledge and agree with the reviewer’s point regarding the model’s current scope and the complexity of industrial crystallization processes. Our present study primarily aimed to validate a novel, data-driven approach for predicting crystal size distributions under laboratory-scale conditions, using relatively simple process variables like seed loading and temperature profiles. We recognize the importance of expanding our framework to include these additional variables.

Also, we added this suggestion in conclusion.

“This framework could also be extended by incorporating additional process variables that play a critical role in other crystallization systems.” Rows 519-520

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript describes a method using  LSTM model to predict particle size evolution behaviors during the nucleation given the temperature profile. This study is significant as it has the potential to replace the current data-processing-intensive method for tracking particle size.  I believe the manuscript should be accepted. However, I do have a few comments:

1. Authors should show typical real-time images collected by IC-CLD, and detail how the image is processed to extract SW.
2. for nucleation control, an interesting concept is that one can interfere during this process to manipulate the nucleation behavior. The study presented here only shows nucleation during natural cooling. I am wondering if the author can test ML model with a non-natural cooling process. such as a sudden temperature jump and drop during the nucleation, and see how effective this model can be in predicting a non-natural cooling process?
3. The authors mentioned in the beginning, that their model included serval features such as dT/dt,d²T/dt, and so on.  I am wondering if the authors could comment on which feature(s) dominate in their trained model.

Author Response

Reviewer Comment:

Authors should show typical real-time images collected by IC-CLD, and detail how the image is processed to extract SW.

Author Response:

We thank the reviewer for this valuable suggestion. In response, we have added a representative microscope image as Figure 3. in the revised manuscript to visually demonstrate the quality and characteristics of the in-situ microscopy data used in this study.

The images are processed using software provided by BlazeMetrics, which performs automatic extraction of chord length distributions (CLD) based on advanced edge detection and particle segmentation algorithms. However, as the exact image processing methodology is proprietary, we do not have full access to the internal algorithmic details.

 

Reviewer Comment:

For nucleation control, an interesting concept is that one can interfere during this process to manipulate the nucleation behavior. The study presented here only shows nucleation during natural cooling. I am wondering if the author can test the ML model with a non-natural cooling process, such as a sudden temperature jump and drop during nucleation, and see how effective this model can be in predicting a non-natural cooling process?

Author Response:

We thank the reviewer for this thoughtful and valuable suggestion. The current study was designed to focus on engineered cooling profiles—both linear and non-linear—to capture a range of crystallization behaviors representative of standard lab-scale process development. While these include dynamic changes in cooling rate (Figure 5. in manuscript), they do not yet cover abrupt thermal disturbances such as sudden temperature jumps or drops during nucleation.

We fully agree that testing the model under such cooling conditions would provide deeper insight into its generalization capabilities and practical robustness. This is an excellent direction for future work. Designing experiments with deliberate thermal shocks during the early stages of crystallization would allow us to assess how well the model can adapt to untrained and possibly out-of-distribution process behaviors.

Reviewer Comment:

The authors mentioned in the beginning that their model included several features such as dT/dt, d²T/dt², and so on. I am wondering if the authors could comment on which feature(s) dominate in their trained model.

Author Response:

We thank the reviewer for this insightful question. In our current work, we incorporated dynamic temperature descriptors such as the first derivative (dT/dt), the second derivative (d²T/dt²), and the cumulative integral of temperature to enrich the feature space and improve the model’s ability to capture the crystallization dynamics.

However, we did not perform a formal feature importance analysis—such as SHapley Additive exPlanations (SHAP)—as part of this feasibility study. The primary aim at this stage was to evaluate whether the proposed machine learning framework could accurately predict crystal size metrics (e.g., d90) under varied cooling strategies.

We fully agree that a rigorous feature sensitivity or importance analysis would add depth to the model interpretability and generalizability.

Reviewer 3 Report

Comments and Suggestions for Authors

The manuscript by Vrban and co-authors introduces a data-driven modeling framework using Long Short-Term Memory (LSTM) neural networks to predict crystal size distribution metrics in seeded cooling crystallization, based on in situ microscopy images. While the study focuses solely on creatine monohydrate and does not explore other APIs, the approach addresses a key challenge in crystallization modeling and avoids the need for direct supersaturation measurements, which are often difficult. In this context, the work has the potential to contribute meaningfully to the field by offering a practical alternative for crystal size prediction. Overall, I found the methodology appropriate, the results convincing, and the topic of interest to readers of Processes.

A few comments:

1. For model building, were confidence intervals or uncertainty estimates included in the model predictions? If not, how do the authors assess the reliability of the results, especially in scenarios involving process scale-up or tighter quality control??
2. Could the authors comment on how other factors like pH, agitation rate, or initial solute concentration might influence model accuracy or be integrated into the current framework??
3. The authors used a MinMax scaler for input/output normalization. Could they clarify why this method was chosen over alternatives like standard scaling (z-score)?

Author Response

1. For model building, were confidence intervals or uncertainty estimates included in the model predictions? If not, how do the authors assess the reliability of the results, especially in scenarios involving process scale-up or tighter quality control?

 

Response:

We thank the reviewer for highlighting this important point. In the current work, the primary focus was on demonstrating the predictive capability of the LSTM model using experimental validation through independent test sets. Confidence intervals or explicit uncertainty quantifications were not directly calculated within the LSTM predictions. However, reliability was implicitly evaluated using robust metrics such as Median Absolute Error (MedAE) and Root Mean Square Error (RMSE) across various independent experimental conditions, including diverse cooling profiles and seed loadings. For practical applications, especially involving tighter quality control scenarios, implementing uncertainty quantification methods would indeed provide additional insights into prediction reliability and variability. Such extensions will be considered in our future work to further enhance the practical applicability and robustness of our modeling approach.

 

2. Could the authors comment on how other factors like pH, agitation rate, or initial solute concentration might influence model accuracy or be integrated into the current framework?

 

Response:

We appreciate this insightful comment. The presented LSTM model framework focused specifically on seed loading and temperature profiles, as these variables were controlled and systematically varied in our experiments. However, other process parameters such as agitation rate, and initial solute concentration indeed have known effects on nucleation and crystal growth dynamics, potentially influencing model accuracy.

Integrating additional process parameters into the LSTM model is straightforward and can be achieved by extending the input feature set to include these variables and their dynamic descriptors. Practically, this would require additional experimental data to systematically characterize the impact of these parameters on crystallization outcomes. Exploring these variables represents an important direction for future research, enhancing the model’s generalizability and applicability to broader crystallization systems.

 

Also, we added this suggestion in conclusion.

“This framework could also be extended by incorporating additional process variables that play a critical role in other crystallization systems.” Rows 519-520

 

3. The authors used a MinMax scaler for input/output normalization. Could they clarify why this method was chosen over alternatives like standard scaling (z-score)?

 

Response:

Thank you for the opportunity to clarify this methodological choice. MinMax scaling was selected primarily due to its simplicity and interpretability, as it transforms all data features into a defined range between 0 and 1. This approach was beneficial for the LSTM network, ensuring numerical stability during training by preventing any single feature from dominating due to magnitude differences.

 

In contrast, z-score (standard) scaling centers features around zero with unit variance, which is often preferable for data following a roughly normal distribution. Given the experimental setup and the practical context of our dataset, where absolute values of temperature and particle metrics have clear physical meanings, MinMax scaling was deemed appropriate. However, it is recognized that standard scaling could also be effective, especially for datasets where normalization around mean and variance might better capture feature variability. Future comparative analyses of different scaling methods could provide further insights into their respective benefits for crystallization modeling.