Spray drying is widely used in pharmaceutical, food, and other process industries to convert liquid feeds into dry powders. Engineers often use CFD simulations to understand temperature distribution and drying behaviour within the dryer. While CFD provides valuable insights, evaluating a large number of operating conditions can become time-consuming.

PhysicsAI models offer an opportunity to accelerate design space exploration by learning from validated CFD simulations and providing rapid predictions of process behaviour. In this context, a PhysicsAI model refers to a machine learning surrogate trained using data generated from validated physics-based simulations.

This article describes a practitioner's journey in developing and validating a PhysicsAI model for temperature field prediction in a spray dryer using the InteliSIM platform. The practitioner need not possess expertise in AI model architectures or coding workflows. In fact, domain expertise is the key requirement for developing, validating, and certifying the PhysicsAI surrogate.

Once trained and validated, the model can be packaged as a process engineer-facing application using InteliSIM. The process engineer need not have expertise in CFD simulations or machine learning to use the application. Process engineers can evaluate operating conditions and visualize temperature fields within seconds without requiring access to CFD software or machine learning workflows.

Figure 1. Practitioner workflow using InteliSIM. A domain expert develops and validates the PhysicsAI model using CFD data and releases it as an application. Process engineers then use the deployed application for rapid design space exploration and process optimization.

ML Model for Spray Dryer

Spray drying is influenced by several interacting variables, and engineers often need to understand their effect on product quality, solvent removal, residence time, and operating efficiency.

Simulation of a spray dryer involves handling complex physics, model validation, and execution of multiple runs to understand the effect of process variables. CFD simulations provide valuable insights into flow patterns, temperature distribution, and drying behaviour inside the chamber. However, evaluating a large number of operating conditions through CFD can become time-consuming.

The goal of the machine learning model is to enable rapid exploration of the design space and help identify optimal operating conditions. The model development process follows a structured approach, and this multi-part series will uncover that journey.

A design space was defined over the following process parameters:

• Drying air mass flow rate
• Drying air inlet temperature
• Spray mass flow rate

Why Temperature Was Selected as the First Milestone

One of the key CFD model validation parameters is the exhaust temperature of the drying air, since it can be measured in an actual setup using a thermocouple. Therefore, prediction of the temperature field was selected as the first milestone for the ML model.

Temperature was selected as the first target variable because it is directly linked to the drying process and can be validated using plant or laboratory measurements. In addition, successful prediction of the temperature field provides a strong foundation for extending the PhysicsAI model toward prediction of other KPIs such as solvent concentration and drying performance in subsequent phases.

The objective was not just to predict the exhaust temperature, but the temperature field throughout the dryer. While the exhaust temperature provides a useful validation point, the temperature distribution within the dryer provides much deeper insight into the process and serves as a stronger test of the model's predictive capability.

The ML predictions are compared against results from a validated CFD model. The CFD model used for generating the training data had already been validated against experimentally measured exhaust temperatures before being used as a source for model development. This ensured that the machine learning model learned from a physically representative process model rather than purely numerical data.

Model Development

As a first step, data from 25 simulations was selected to train the model. The simulations covered operating conditions within the selected design space and provided the temperature field data required for training.

Once trained, the ML model required less than 2 minutes to predict the complete temperature field throughout the dryer.

Model Validation

The model accuracy was compared against CFD results by comparing the CFD and ML predictions at every cell of the CFD mesh rather than using an average error metric. This represents a significantly stricter validation criterion, since local deviations anywhere within the dryer contribute to the error.

For a temperature field containing thousands of computational cells, this approach evaluates the ability of the model to reproduce local temperature variations throughout the domain rather than simply matching a global average value.

The image below shows this comparison, and the results were quite satisfactory. Across the evaluated cases within the design space, the cell-wise error remained well within 5%.

Figure 2. Comparison of CFD-predicted and ML-predicted temperature fields, along with the corresponding error contour, for a representative operating condition within the training design space.

Generating the same temperature field through CFD requires several hours of simulation time, whereas the trained PhysicsAI model generates the complete temperature field in less than 2 minutes.
As a next step, the model accuracy was assessed at various points both within and outside the design space.
As shown in the image below, the model accuracy within the design space remains well within 5%.

Figure 3. Model accuracy assessment at multiple operating conditions within the training design space.

Case 4 above is, in fact, marginally outside the design space and still provides a reasonably acceptable prediction. However, as expected, the model accuracy degrades as operating conditions move further away from the training design space.

This behaviour is consistent with most machine learning models and highlights the importance of defining an appropriate design space during model development. The model is intended for rapid prediction within the trained operating envelope and is not a replacement for CFD when evaluating completely new operating regions.

Figure 4. Model performance for conditions outside the training design space. Accuracy remains acceptable close to the design space boundary but gradually degrades as operating conditions move further away from the training envelope.

Deployment as a Practitioner Application

Once validated, the PhysicsAI model can be packaged and released as a practitioner-facing application using InteliSIM. The application allows process engineers to evaluate operating conditions and visualize predicted temperature fields without requiring access to CFD software or machine learning workflows.

This approach enables engineering knowledge captured through simulations to be validated, packaged, and made accessible to a wider group of users within the organization.

Figure 5. InteliSIM application generated and released by the practitioner for use by process engineers.

Takeaways

• A PhysicsAI model capable of predicting the complete temperature field within minutes offers significant value in accelerating process innovation and process optimization studies.
• The InteliSIM platform enabled the practitioner to set up the model training workflow in less than 10 minutes.
• Model accuracy is better within the design space than outside it, as expected.
• InteliSIM enabled model training on an entry-level GPU workstation typically used by a practitioner.
• The model is intended for rapid prediction within the trained operating envelope and complements, rather than replaces, detailed CFD simulations.
• Validation of the temperature profile was the first step toward building this model. The next phase extends this work to predict solvent concentration within the dryer. While temperature governs the driving force for drying, solvent concentration is more directly related to drying performance and final product quality. Accurate prediction of solvent concentration would enable faster assessment of drying efficiency and process optimization opportunities.
• This series will progressively cover temperature prediction, solvent concentration prediction, and eventual deployment of practitioner-facing applications using InteliSIM.