AI Takes the Wheel in the Quest to Mass-Produce Atomically Thin Materials

Two-dimensional materials promise to reshape modern technology. Graphene, transition metal dichalcogenides, hexagonal boron nitride, and their relatives exhibit extraordinary electronic, optical, and mechanical properties that arise from their sheer thinness—a single layer of atoms. Yet the dream of integrating these materials into commercial chips, sensors, and flexible electronics has been hampered by a stubborn problem: making them consistently, at scale, and to specification remains extraordinarily difficult. A new perspective published in npj Advanced Manufacturing argues that the solution lies not in better furnaces alone, but in marrying chemical vapor deposition—the workhorse technique for growing 2D films—with artificial intelligence and autonomous laboratory systems.

Chemical vapor deposition, or CVD, is conceptually simple. Precursor gases or vapors are introduced into a heated chamber where they react on a substrate surface, nucleating and growing crystalline films layer by layer. In practice, however, the process is a nightmare of coupled variables. Temperature gradients across the furnace, gas flow rates, precursor partial pressures, substrate surface chemistry, chamber geometry, cooling ramps, and even trace contaminants all influence whether the growth yields a pristine monolayer, a useless patchwork of multilayer islands, or nothing at all. Small changes in one parameter can ripple through the entire system in nonlinear ways, which is why transferring a growth recipe from one laboratory’s furnace to another’s often fails spectacularly.

Traditionally, researchers have navigated this parameter space the way navigators charted unknown seas: iteratively, slowly, and guided by intuition. A doctoral student might spend months running growth after growth, characterizing each sample under a microscope or Raman spectrometer, and adjusting conditions based on experience. The authors of the new analysis contend that this manual, serial approach is fundamentally mismatched to the complexity of 2D material synthesis. The parameter space is too large, the interactions between variables too intricate, and the experimental throughput too low for human trial-and-error to deliver the reproducibility that industrial manufacturing demands.

Enter machine learning. By training models on existing growth data—including published literature and internal laboratory records—researchers can build surrogate models that predict outcomes such as domain size, layer number, crystal orientation, and defect density from synthesis parameters without running an experiment. Bayesian optimization algorithms then use these surrogates to select the next experiment intelligently, balancing exploration of uncertain regions of parameter space against exploitation of promising conditions. Instead of one hundred blind experiments, a well-designed closed-loop campaign may converge on an optimal recipe in a few dozen. The perspective highlights that such approaches have already accelerated optimization of graphene growth on copper foils, uniform monolayer MoS2 deposition, and the control of grain boundaries and twist angles between stacked layers.

But predictive models are only half of the equation. The other half is quality control—a persistent bottleneck in moving 2D materials from laboratory curiosities to certified manufacturing feedstock. Conventional characterization, relying on Raman spectroscopy, photoluminescence mapping, atomic force microscopy, and electron microscopy, is slow, sample-intensive, and often destructive or off-line. The authors argue that AI can transform this step as well. Deep learning models trained on optical microscopy images can classify layer number, identify grain boundaries, and flag defects across wafer-scale samples in minutes. Spectral analysis algorithms can extract quantitative information from Raman and photoluminescence data with greater consistency than human operators. Integrating these inspection tools directly into the synthesis workflow enables real-time feedback: if the model detects that the growing film is drifting out of specification, the process parameters can be adjusted mid-run rather than after the fact.

This convergence of synthesis, sensing, and machine intelligence points toward the article’s central vision: the self-driving laboratory. In an autonomous manufacturing platform, robotic systems handle substrate loading, precursor delivery, and sample transfer; in-line sensors stream data on film growth and quality; and a decision-making algorithm orchestrates the entire loop, proposing experiments, executing them, evaluating results, and refining its strategy. Humans set the goals—the target material, desired thickness, crystal quality, and throughput constraints—while the automated system navigates the path toward them. Early demonstrations of such platforms in materials chemistry, including autonomous flow chemistry rigs and robot-driven thin-film laboratories, suggest the concept is not science fiction but an emerging engineering reality.

The implications for the semiconductor industry are considerable. As device architectures push into the angstrom scale, silicon itself is running out of room to shrink, and 2D materials are among the leading candidates for next-generation transistors, memristors, photodetectors, and quantum devices. The International Roadmap for Devices and Systems has identified monolayer transition metal dichalcogenides as potential channel materials for future logic technology. But fabs operate on tolerances measured in fractions of a percent, and no conventional CVD process today delivers that level of wafer-to-wafer uniformity at scale. AI-driven process control, with closed-loop adjustment based on in-line metrology, offers a credible route to closing that gap—provided the underlying data infrastructure is built to industrial standards.

The perspective is candid about the obstacles. Machine learning in materials synthesis is data-hungry, yet most growth data in the literature is sparse, heterogeneous, and poorly documented. Negative results—failed growths—are rarely published, biasing datasets toward successes and limiting the ability of models to learn failure boundaries. Different laboratories report parameters inconsistently, and furnace-to-furnace variation means that even well-documented recipes are not directly transferable. The authors call for standardized data reporting formats, shared repositories of synthesis records, and the adoption of emerging frameworks such as FAIR data principles—findable, accessible, interoperable, reusable—so that community-scale datasets can be assembled and models can generalize beyond a single instrument.

Safety, reproducibility, and interpretability also demand attention. An autonomous system running hazardous precursors such as metal-organic compounds, hydrogen sulfide, or hydrogen at high temperatures must incorporate robust safeguards, fault detection, and emergency protocols. And for AI recommendations to be trusted by process engineers, the models must not be black boxes; explainable machine learning approaches that reveal which parameters drive a given prediction will be essential for regulatory compliance and for building human confidence in automated decision-making. The authors envision hybrid workflows in which human experts and autonomous systems collaborate, with the AI handling routine optimization while scientists tackle genuinely novel materials discovery.

If those challenges are met, the payoff extends well beyond any single material. A mature, AI-enabled CVD infrastructure would function as a general-purpose platform: the same closed-loop hardware and software stack could be retargeted from graphene to MoS2 to emerging quantum materials simply by swapping precursors and retraining the models. That flexibility, combined with dramatically reduced development timelines, could finally push 2D materials out of the cleanroom demonstration phase and into mass production—powering faster transistors, ultrasensitive sensors, flexible displays, and photonic circuits that today exist only on whiteboards. The message of the new analysis is clear: the bottleneck in 2D materials manufacturing is no longer a lack of ideas but a lack of control, and artificial intelligence is poised to supply it.

Subject of Research: AI-enabled chemical vapor deposition synthesis, quality control, and autonomous manufacturing of two-dimensional materials

Article Title: AI-enabled CVD synthesis, quality control, and autonomous manufacturing of 2D materials

Article References: Mannan, M. I., Mistry, P., Kansal, V., Estrada, V., & Leem, J. (2026). AI-enabled CVD synthesis, quality control, and autonomous manufacturing of 2D materials. npj Advanced Manufacturing. https://doi.org/10.1038/s44334-026-00109-5

Image Credits: AI Generated

DOI: 10.1038/s44334-026-00109-5

Keywords: two-dimensional materials, chemical vapor deposition, machine learning, autonomous laboratories, graphene, MoS2, transition metal dichalcogenides, quality control, semiconductor manufacturing, Bayesian optimization, self-driving laboratory, advanced manufacturing