DescriptionDigitization of manufacturing processes has led to an increase in availability of process data which has enabled the use of data-driven models to predict the outcomes of these manufacturing process. Data-driven models are instantaneous to simulate and can provide real-time predictions but lack any governing physics within their framework. When process data deviates from original conditions, the predictions from these models may not agree with physical boundaries. In such cases, the use of first principle-based models to predict process outcomes have proven to be effective but computationally inefficient. Data-driven model unlike their first-principle counterparts do not have the ability to scale across geometries. Thus, there remains a need to develop computationally efficient models with physical understanding about the process. Training data provided governs the range of the prediction of the data-driven model, thus these models also need for these models to adapt to new data, outside of the original domain is provided.
In the first case study, the addition of physics-based boundary conditions to a neural network to improve its predictability for granule density and granule size distribution (GSD) for high shear granulation process was demonstrated. The physics-constrained neural network (PCNN) trained slower than traditional neural networks but with constrained outputs, it was able to predict the granule size growth within the steady growth regime accurately. When input data which violated physics-based boundaries was provided, the outputs from PCNN identified these points more accurately compared to other non-physics constrained neural network with an error of <1%. Thus, after incorporating physics, scaling of the model across geometries and different materials was undertaken. For this study, a collection of already published experimental data for twin screw granulation as leveraged. Several input parameters and outputs quality attributes were collected. Missing outputs were completed using data imputation which included both first principle as well as statistical methods. Completion of this data set also led to understanding the process dynamics and granule growth. This knowledge was then utilized to enforce physical boundaries on the outputs of a supervised autoencoder model. An added advantage to using an autoencoder over a traditional neural network was determination of reduced order latent variables that can describe the process dynamics well. In the final case study, the aim is to utilize the earlier developed steady state model in a more dynamic environment. The model would simultaneously optimize the energy efficiency of the twin screw granulation process while retraining the model itself when unknown conditions are encountered. This study would give the physics-constrained model the ability to extend its range for more accurate predictions and decreasing energy consumption.