| dc.description.abstract | Real time monitoring and control of crystallization processes in process industry producing food, fine chemicals and pharmaceuticals is of major importance to assess quality and purity of the product. The lack of available online and in-situ measurement options have been identified as a bottleneck in this regard. The Focused Beam Reflectance Measurement (FBRM) probe can access online and in-situ measurements of the chord length distribution (CLD). However, this measurement only provides information related to the particle size, and thus a conversion to the particle size distribution (PSD) is needed. This problem has been subject to research in literature. There are several ways to address this problem, either by mapping PSDs to CLDs called the forward problem or the mapping of CLDs to PSD called the inverse problem. In this research, a framework to solve the forward problem is presented. The suggested framework consists of two main structures. The first part is a convolutional neural network (CNN) which maps PSDs to CLDs. The CNN is trained with in-silico generated data. This CNN model has already been developed from previous research efforts by members of the Mesbah research group at University of California, Berkeley. In this research, a second structure has been developed. It consists of a dimensionality reduction and correction layer to adapt the framework to an experimental dataset in the low data regime. It accounts for possible discrepancies between the experimental and simulated dataset and considers experimental conditions. The dimensionality reduction was performed with an autoencoder by reducing the dimensionality of the CLDs. The correction layer was modeled with a multi-output Gaussian process regression model.
The autoencoder was implemented and displayed great ability to compress and reconstruct CLDs from both the simulated and experimental dataset. With this tool, the CNN model was adapted and trained with the reduced dimensionality data. The new CNN model exceeded the previous model with regards to accurately predicting CLDs with the simulated dataset, however it failed to predict CLDs from the experimental data set. With the implementation of the correction layer, the final framework was successfully able to map the experimental PSDs to CLDs. This developed framework demonstrates how both simulated and experimental data can be leveraged to develop an accurate data-driven model for prediction without requiring substantial amount of experimental data. This framework also provides a robust model to generate enough data to address the inverse problem (CLD to PSD) which is of greater interest. | |