Data-driven graphene synthesis and additive manufacturing

Shah, Aagam Rajeev

Permalink

https://hdl.handle.net/2142/129597 Copy

Description

Title

Data-driven graphene synthesis and additive manufacturing

Author(s)

Shah, Aagam Rajeev

Issue Date

2025-04-29

Director of Research (if dissertation) or Advisor (if thesis)

• Tawfick, Sameh
• Ertekin, Elif

Doctoral Committee Chair(s)

van der Zande, Arend

Committee Member(s)

• Shoemaker, Daniel P
• Charpagne, Marie A

Department of Study

Materials Science & Engineerng

Discipline

Materials Science & Engr

Degree Granting Institution

University of Illinois Urbana-Champaign

Degree Name

Ph.D.

Degree Level

Dissertation

Keyword(s)

• bayesian optimisation
• graphene
• chemical vapour deposition
• laser powder bed fusion
• computer vision.

Abstract

The increasing complexity of synthesis and manufacturing processes poses significant challenges to their optimisation. Traditional physics-based models struggle to keep pace with the rapid development of these processes, often constrained by the computational time and power required to accurately model systems across multiple time and length scales. My doctoral research addresses these limitations by demonstrating the use of data-driven techniques to optimise experimental parameters with computationally inexpensive surrogate models. Specifically, I focus on two systems that stand to benefit from such an approach: graphene synthesis via chemical vapour deposition (CVD) and metal additive manufacturing using laser powder bed fusion (LPBF). Graphene synthesis via chemical vapour deposition (CVD) – a process where hydrocarbon precursor gases decompose on metal catalysts at elevated temperatures to form atomically thin carbon layers – has emerged as a reliable method for growing large area graphene. However, the process remains highly sensitive to many of its ∼300 dependant variables such as temperature, gas flow ratios, steps sequence and durations, oven design, and environmental conditions. This sensitivity is difficult to quantify because the large number of control parameters affect a small number of quantifiable graphene properties such as the growth area, throughput, and graphene quality. The high sensitivity persistently limits predictability and repeatability. This inherent variability makes it extraordinarily difficult to identify optimal recipes that balance growth quality, domain size, and throughput. First-principles modelling has provided limited progress in addressing these challenges because they cannot capture the full timescale of growth and the important macroscale effects of the oven design, gas flow dynamics, and slow time transport phenomena in a large reactor. To overcome this, I propose a three-step workflow: (1) systematic data collection guided by design of experiments principles, (2) development and training of a neural network to automatically analyse scanning electron microscopy (SEM) images of graphene and extract quantitative metrics, and (3) application of a Gaussian process model coupled with Bayesian optimization to refine experimental parameters. I demonstrate the efficacy of this method by maximizing graphene coverage as an initial proof-of-concept. Furthermore, I extend the analysis to model growth rate and domain size, offering valuable insights into the growth dynamics of graphene. Our model shows the dependence of domain size and growth rate on the growth time and the ratio of the partial pressures of CH4 and H2. We find that the domain size decreases from a peak with increasing partial pressure of H2, and increases with an increase in the growth time. Our model shows that the instantaneous areal growth rate of graphene increases to a peak and then decreases over time at constant partial pressures of H2. This peak gets delayed as we increase the partial pressure of H2, leading us to a hypothesis of a hydrogen-dependent mechanism of graphene growth. Metal additive manufacturing via LPBF is a layer-wise process where a high-power laser selectively melts fine metal powder particles to build complex three-dimensional components. It has become indispensable for aerospace, medical and automotive applications requiring near-net-shape parts with intricate geometries. Despite its versatility, the performance of LPBF is governed by nonlinear interactions between dozens of parameters, including the laser power, scan speed, powder size distribution, and material properties. These interdependencies create process instabilities that manifest as defects like porosity, residual stress, and microstructure inconsistencies, mirroring the challenges seen in the CVD synthesis of graphene and complicating efforts to establish robust parameter sets. Applying the same data-driven workflow, I leverage experimental data generated by collaborators to design and train a neural network capable of automatically extracting melt pool geometry from optical microscopy images of the cross sections of single metal tracks. Non-dimensional scaling laws - which collapse process parameters like energy density and cooling rates into universal invariants - have emerged as critical tools for predicting system performance. I applied the three-step workflow to LPBF. My preliminary results highlight the potential for active learning models to leverage these non-dimensional parameters to further enhance material-agnostic parameter optimisation and deepen our understanding of LPBF processes. This research underscores the transformative potential of data-driven methodologies in overcoming the limitations of traditional approaches, paving the way for more efficient and scalable optimisation in advanced synthesis and manufacturing systems.

Graduation Semester

2025-05

Type of Resource

Thesis

Handle URL

https://hdl.handle.net/2142/129597

Copyright and License Information

Copyright 2025 Aagam Rajeev Shah

Owning Collections