R. Sappington, B. Cornick
Epic Advanced Materials,
Keywords: machine vision, boron nitride nanotubes, synthesis, convolutional neural networks
Summary:Despite significant advances in recent years in the fields of nanotechnology and nanomaterials, industry adoption and commercialization of nano-enabled technologies has been slow. Nanomaterials have been studied in research labs for decades now, and yet it is still difficult to find a commercially available product that demonstrates their benefits as proven on a laboratory scale. This is primarily due to the low availability and relative high cost of most nanomaterials, caused by a bottleneck in production capacity and a variable consistency in nanomaterial quality. Nanomaterial synthesis processes have historically been difficult to scale up to the level needed to eliminate this bottleneck. This can be attributed to process inefficiencies and sub-optimal product quality that is inherent in large batch nanomaterial production processes. Further complicating the issue is the vast, and often unexplored, parameter spaces that these processes can live in. New technologies at the intersection of machine learning (ML) and advanced manufacturing are emerging in an effort to overcome these barriers to commercial scale nanomaterial synthesis and their subsequent adoption into industry. In this work, Epic Advanced Materials investigates the use of automated systems and machine learning algorithms to assist in the transition from small scale batch to commercial scale continuous or semi-continuous nanomaterial synthesis processes. The Extended Pressure Inductively Coupled (EPIC) synthesis process, which can be used in the production of boron nitride nanotubes (BNNTs), is presented as a case study to demonstrate the efficacy with which ML algorithms can predict and optimize the parameters that govern nanomaterial nucleation and growth in a continuous or semi-continuous flow reactor. Optimizing the process for desirable material properties by traditional approaches has proven to be challenging and requires the use of multi-dimensional design-of-experiment (DOE) analysis, extensive manual identification and calculation of parameter interactions, and large amounts of costly experimental data. In this work we discuss the key process control parameters and the identification of important parameter interactions through a simplified design-of-experiments approach. Our approach then draws upon algorithms: VGG, Inception, ResNet and U-Net architectures to explore the most optimal convolutional neural nets (CNNs) for image classification of BNNTs. Such model exploration will involve feature generation and ranking to arrive at highest priority and representative features of BNNT purity and impurity. Such model exploration and feature engineering incorporates transfer learning to take advantage of model performance and accuracy. Data selection and sharing is then further explored in the context of commercial research and development, focusing on increasing the availability of training data while also addressing concerns of intellectual property protection. While the initial focus of this work covers a specific process for the commercial synthesis of BNNTs, the ability to quickly explore vast manufacturing process parameter spaces and precisely control the key process parameters will allow for a quick optimization of manufacturing procedures that can be readily extended to other material systems.