Matt Maschmann, Prasad Calyam, Filiz Bunyak, Kannappan Palaniappan

Materials Problem

Thirty years after their discovery, carbon nanotube (CNT)-based products and devices are limited, despite their superior mechanical, thermal, and electrical properties. The technological bottleneck lies not strictly in the material itself, but in its reliable and scalable synthesis. The chemical vapor deposition (CVD) parameter space to synthesize CNTs is vast and multi-dimensional, including catalyst composition, catalyst thickness, substrate composition, synthesis temperature, synthesis pressure, gas composition, among numerous others. The combinations of these parameters are inexhaustible, and researchers have only sparsely explored the available parameter space – largely because of a lack of physical insight into the underlying mechanisms.

Dense CNT arrays, typically comprised of 108-1010 CNT/cm2, are an attractive option for scalable CNT production. The physical properties of CNT arrays, however, are severely diminished relative to those of individual CNTs, because of the cellular, interwoven structural morphology of CNT forests developed during their synthesis and self-assembly (Figure 8b).

The reported stiffness of CNT arrays, for example, varies by orders of magnitude [32–35], but is always significantly less than the volumetrically scaled modulus of an isolated CNT. If the synthesis-structure-property relationships of CNT arrays could be determined and controlled, CNT forests with diverse physical property sets may be achievable.

This project integrates in-situ scanning electron microscope (SEM) synthesis, finite-element simulation, machine learning, and artificial intelligence to determine these correlations for morphology CNT arrays. The application domain of interest is piezoelectric sensitivity for the measurement of small dynamic loads and displacements [33,36]. The findings may also be extended to CNT-based systems such as multifunctional composites, interfaces, and optical absorbers.

Data Science Approach

This project uses a cyber-physical research system (Figure 8) to autonomously explore the CVD synthesis parameter space of CNT arrays, with the goal of deterministically synthesizing CNT forests with diverse property sets. As an example, we wish to demonstrate CNT forests with compressive moduli that vary by at least four orders of magnitude. The synthesis system will use both physical and numerical experiments to accomplish this goal. The main components of the system include i) in-situ SEM observation of CNT forest synthesis, ii) in-line image analysis of catalyst particle size, CNT growth rates, and CNT-CNT interactions, iii) a physics-based simulation to increase data throughput and exploration of new parameters [37–40], iv) deep neural networks to identify synthesis-structure-property relationships [41], v) indentation to confirm properties and vi) distributed control between human and cyber decision makers.

The machine learning techniques will utilize real-time analysis of SEM images, physics-based simulation, and post-synthesis electromechanical evaluation as input data streams. The system will use AI concepts such as expert systems, knowledge management, and case-based reasoning to launch new in-situ SEM syntheses and corresponding simulations until the desired electromechanical response is obtained.

Physical in-situ CNT array synthesis experiments will generate time-resolved image sequences that will be analyzed with computer vision. A physics-based synthesis simulation (Figure 9e) will complement physical experiments. Unique simulations may be run in parallel using high performance computing resources, facilitating rapid data collection. Deep neural networks will then be used to correlate the CNT array morphology with mechanical and electrical properties obtained using nanoindentation [33,36]. Distributed experimental control will evolve with time as the AI becomes more competent. Human researchers would first control experiments to train AI. Over time, the AI will decide synthesis parameters and autonomously launch and control experiments without human intervention. To support the parallel experimentation, Dr. Scott (Co-PI) will facilitate deployment of research code to the NSF Nautilus cluster (Pacific Research Platform), as he has with a variety other research projects in his role as the PI of an NSF Regional CyberTeam.

Opportunities for Creativity

Avenues for creative thought are presented for each of the physical and numerical systems employed in this project. The vast range of available CNT array properties suggests that diverse applications can be imagined. To enrich discussions about applications, analogies to fibrous systems in biology will be sought. With the physical and numerical techniques, previously assumed constraints will be challenged and then re-imagined. Optimization of the finite element simulation, perhaps guided by data science, is another area in which creativity thought could generate massive performance gains.