AIM2D White paper

Background. Recent breakthroughs in two dimensional (2D) materials and their heterostructures, especially moiré quantum matter, have led to the discovery of exotic quantum phases, such as sliding ferroelectricity, superconductivity, and even fractional electronic phases. These new phases have the potential to significantly boost quantum applications. The canonical example is twisted bilayer graphene: at the “magic” angle ~1.1°, flat electronic bands amplify interactions, producing unconventional superconductivity correlated insulating states, and a suite of topological phases

Challenge**.**** **The current paradigm for fabricating 2D devices is fundamentally manual, suffering from low throughput and poor reproducibility. Fabrication rates are typically limited to a single device per week per expert, while key parameters like yield, cleanliness, and twist-angle accuracy vary significantly. This inconsistency stems from uncontrolled variables and a lack of standardized, machine-readable process data. Although emerging computer-vision tools can identify materials[1-3], they are not integrated into a closed-loop manufacturing system; they cannot dynamically adapt fabrication protocols, perform statistical optimization, or correlate process parameters with final device performance. This creates a critical bottleneck, widening the gap between laboratory discovery and scalable technology. This bottleneck extends beyond fabrication into device characterization. The scientific discovery process for 2D devices requires the exploration of a high-dimensional parameter space with many tunable degrees of freedom, such as electrostatic gating, temperature, magnetic field, and current bias. The current reliance on manual exploration of this complex space is inefficient and has significantly slowed the discovery of novel quantum phases. An end-to-end ecosystem that uses AI to automate and integrate both fabrication and characterization is therefore essential to accelerate scientific progress.

AI**-augmented High-throughput Manufacturing of Two-Dimensional Devices (AIM2D)**. We propose AIM2D, an AI-supervised robotic platform capable of fully automated and scalable fabrication and characterization of 2D devices. The system integrates four key layers, including oversight, cognitive, control, metrology & safety and physical layer as shown in Fig. 1.

**Technical details. **** **

1. The oversight layer: Human-AI collaboration. This highest layer ensures that the entire autonomous system remains aligned with the research goals and under supervision.** ** We have a dashboard that provides monitoring and alerts, and roles like principal investigators (PIs) who set the high-level scientific goals and provide authorization for critical steps, ensuring a human-in-the loop architecture. We fully unlock the interdisciplinary power of AI by having AI experts & consultants: the system can leverage specialized AI models (e.g., fine-tuned or prompted for polymer chemistry or electrical engineering) to consult on complex, domain-specific problems.

2. The system's intelligence is powered by the LabAgent Framework, an AI orchestrator responsible for all high-level decision-making, planning, and optimization. Its core functions are: Workflow Planner: Uses Retrieval-Augmented Generation (RAG)[4] and a persistent knowledge base to autonomously generate and sequence the steps for device fabrication. Process Optimizer: Dynamically tunes experimental parameters, such as temperature and robotic movement speed, to maximize success rates and device quality. Multimodal Reasoner: Analyzes and interprets diverse data types (e.g., microscope images, spectra) to make sophisticated judgments, like assessing the quality of a 2D material flake. Device Diagnostic: Employs a ReAct (Reasoning and Acting)[5] framework to perform final quality control, diagnose fabrication flaws from measurement data, and propose corrective actions for subsequent runs, enabling continuous improvement.

Figure 1. Flow chart of the proposed AIM-2D platform.

3. The Control & Safety Layer. This layer functions as a standardized and safe interface for all physical hardware, bridging the AI with the lab equipment. Modules: The layer is divided into FabMCP for controlling fabrication tools (robotics, stages) and InstrMCP for managing measurement instruments (spectrometers, AFMs, transport measurements). Protocol (MCP)[6]: Each module uses a structured Model Context Protocol. The AI sends a command following a specific template. The MCP then validates this command against safety rules, executes it on the physical hardware, and returns the results in a standardized format. This ensures robust, safe, and reliable operation without requiring the AI to manage complex hardware drivers.

4. The Physical Layer: Hardware & Material Flow. This layer contains the robotic hardware that handles all physical tasks, moving materials from their raw form to a completed device, as detailed in Fig. 2a. It operates in three main stages: I. Preparation: The system sources 2D materials through two automated methods: using large-area films from chemical vapor deposition (CVD) or exfoliating flakes from bulk crystals with a multi-roller system. An AI vision module then inspects and selects the highest-quality flakes (Fig. 2b). II. Stacking: A robotic arm, guided by the AI, precisely assembles the 2D materials into heterostructures. A fully motorized and heated stage allows for automated cutting, sub-micron positioning, and exact twist-angle alignment between layers. The system can perform quality checks mid-assembly without losing alignment (Fig. 2c). III. Characterization: Finished structures are automatically analyzed using an integrated Atomic Force Microscope (AFM) and Raman spectrometer.

Several innovative design elements distinguish our platform from previous robotic fabrication systems (Fig. 2d). The robotic arm features a modular tool-changer that supports multiple end-effectors, including a chip-tray gripper, a stamp manipulator, and a silicon-chip gripper. This flexible toolset allows seamless switching between tasks and paves the way for future expansion. By coupling these programmable capabilities with AI reasoning, the system can intelligently select the optimal tools and strategies for each step. We believe that this freely programmable, semi-modular robotic platform represents a transformative advance in 2D material device fabrication—enabling continuous integration of new functionalities and unlocking the full potential of AI-driven manufacturing.