Gallium nitride (GaN) epitaxy is entering a new phase of semiconductor manufacturing. As demand grows for high-performance RF devices and power electronics, manufacturers are placing greater emphasis on epitaxial quality, process stability, and wafer-to-wafer consistency.
Traditionally, GaN epitaxy has relied on carefully developed growth recipes, experienced process engineers, and continuous monitoring of reactor conditions. However, epitaxial growth involves numerous interacting variables. Even minor changes in substrate temperature, chamber pressure, material flux, or surface reconstruction can influence the resulting crystal structure and electrical properties.
Artificial intelligence (AI) is opening up a new approach. By combining in-situ Reflection High-Energy Electron Diffraction (RHEED), equipment sensor data, and machine-learning algorithms, manufacturers can move beyond fixed growth recipes toward more adaptive, data-driven process control.
In September 2026, South Korea's IVWorks reported an AI-enabled GaN epitaxy manufacturing system that analyzes RHEED data alongside molecular beam epitaxy (MBE) equipment information. According to the company, its accumulated datasets include more than 330 million time-series process data points and over 30 TB of RHEED data collected since 2019.
The development illustrates a broader industry trend: semiconductor epitaxy is gradually moving from recipe-based automation toward increasingly intelligent and potentially autonomous manufacturing.
GaN is widely used in RF communications, high-frequency electronics, and power semiconductor applications. In these devices, epitaxial quality directly affects electrical performance, reliability, and manufacturing yield.
Important GaN epi-wafer characteristics include:
Epitaxial layer thickness and uniformity
Surface roughness and morphology
Crystal defect density
Interface quality
Carrier concentration and electron mobility
Sheet resistance and its uniformity
Buffer leakage characteristics
Wafer stress, bow, and warp
Particle contamination
Achieving the required properties involves controlling several interacting growth parameters, including substrate temperature, gallium and nitrogen fluxes, growth rate, chamber pressure, plasma conditions, doping flux, growth time, and surface reconstruction.
These parameters cannot always be optimized independently. A temperature change, for example, may affect surface diffusion, material incorporation, and growth kinetics simultaneously. The resulting variation may not be immediately visible at the equipment level, yet it can influence the final wafer's electrical or structural characteristics.
This complexity makes real-time process monitoring and early detection of abnormal growth conditions particularly valuable.
Reflection High-Energy Electron Diffraction is an established in-situ characterization technique used in MBE.
During growth, a high-energy electron beam strikes the wafer surface at a shallow angle. The resulting diffraction pattern provides information about the crystal surface without requiring the wafer to be removed from the reactor.
Depending on the growth conditions, RHEED patterns can reveal changes associated with:
Surface reconstruction and crystalline order
Growth mode and surface morphology
Surface roughening
Layer formation and growth transitions
For example, relatively smooth two-dimensional growth can produce characteristic diffraction streaks. As the surface becomes rougher or three-dimensional islands develop, the pattern may change.
RHEED is therefore useful not only for observing the growth process but also for identifying changes that may indicate a transition away from the intended growth regime.
However, interpreting a continuous stream of diffraction images presents a challenge. Conventional visual inspection depends heavily on operator experience, and subtle changes may be difficult to quantify consistently across long growth runs.
Machine learning offers a way to analyze RHEED data systematically and at scale.
Instead of relying exclusively on human interpretation of individual images, an AI model can evaluate image sequences and identify changes in diffraction intensity, streak width, spot formation, pattern geometry, and temporal evolution.
A study published in the Journal of Vacuum Science & Technology A demonstrated machine-learning analysis of RHEED images during MBE growth to identify changes in deposition regimes in real time. Its analysis workflow was designed to operate in less than one second, illustrating the potential for rapid diffraction analysis within a process-monitoring system.
The key advantage is the ability to convert complex image sequences into measurable indicators of growth behavior.
When these indicators are combined with reactor sensor data, the system can examine relationships between observed surface conditions and operating parameters. This creates opportunities to detect abnormal growth states earlier and compare current growth behavior with historical process data.
Nevertheless, AI predictions depend on the quality and representativeness of the training data. A model developed for one material system or reactor configuration cannot automatically be assumed to perform equally well in another.
AI-assisted RHEED analysis becomes particularly valuable when it is integrated into a feedback-control architecture.
A simplified closed-loop system follows this sequence:
A camera captures RHEED diffraction patterns during growth.
Image-processing software extracts relevant features.
A machine-learning model evaluates the current growth state.
The control system determines whether the process is operating within the intended window.
Where validated control rules permit, selected MBE parameters are adjusted.
Subsequent RHEED data are analyzed to evaluate the response.
This approach differs from conventional automation, which primarily executes predefined instructions. A feedback-enabled system can use current process information to inform subsequent decisions.
For example, a model may identify diffraction changes associated with a transition from smooth layer-by-layer growth toward a less desirable growth mode. A supervisory controller could then assess whether an adjustment to substrate temperature or material flux is appropriate.
Such adjustments must remain within a validated process window. AI-based monitoring does not eliminate the need for physical process models, engineering constraints, or safeguards against unstable control actions.
The longer-term objective is increasingly autonomous epitaxy, in which monitoring, decision-making, and process adjustment are integrated into a coordinated manufacturing system.
Uniformity is a critical requirement for commercial GaN epi wafers, particularly as manufacturers move toward larger wafer formats.
Depending on the device structure, customers may evaluate thickness uniformity, sheet resistance, carrier concentration, mobility, aluminum composition, surface morphology, and wafer bow.
AI-assisted process control may contribute to better uniformity in several ways.
MBE reactor conditions can change over time because of source depletion, chamber coating, equipment aging, temperature-calibration changes, and plasma-source variation.
By analyzing RHEED patterns alongside temperature, pressure, and other equipment data, an AI system may identify changes in process behavior before they become obvious in conventional monitoring.
Changes in diffraction patterns can provide early indications that the surface is entering an undesirable growth regime. Timely detection allows engineers to investigate the cause and, where appropriate, intervene before the deviation becomes more severe.
Historical process records can be used to compare current growth trajectories with previously characterized runs. This may help identify deviations associated with changes in electrical properties, surface morphology, or other wafer-quality indicators.
These benefits remain dependent on validation. Improved process monitoring does not automatically guarantee better thickness uniformity, lower defect density, or higher production yield; those outcomes must be confirmed through post-growth measurements.
An effective AI process-control platform requires more than a large collection of sensor readings.
Useful datasets may include RHEED videos, substrate temperature, chamber pressure, material flux, plasma power, source temperature, growth rate, growth time, and maintenance history.
These records become more valuable when linked to post-growth characterization, including:
X-ray diffraction (XRD)
Atomic force microscopy (AFM)
Hall-effect measurements
Sheet-resistance mapping
Photoluminescence
Surface inspection
Epitaxial thickness mapping
The objective is to establish a reliable relationship between process conditions, in-situ surface behavior, and final wafer properties.
For example, if a particular combination of diffraction-pattern evolution and reactor conditions repeatedly correlates with an unfavorable post-growth result, the model may learn to flag similar conditions in subsequent runs.
However, correlation alone does not prove causation. Process engineers must verify whether the identified relationships are physically meaningful and reproducible before using them to guide production decisions.
GaN epitaxial defects can include threading dislocations, surface pits, hillocks, cracks, particles, interface defects, local roughness, and non-uniform doping.
Many defects are influenced by substrate quality, lattice mismatch, thermal stress, buffer-layer design, and other fundamental materials considerations. AI cannot eliminate these physical limitations.
Its more immediate opportunity is to help control defects associated with process instability.
By identifying recurring patterns in in-situ measurements and comparing them with post-growth inspection results, an AI platform may help manufacturers recognize process conditions associated with elevated defect risk.
This can support earlier intervention, more targeted troubleshooting, and improved process-window management.
The practical objective is not to replace materials science, but to maintain growth conditions within a narrower and more repeatable operating range.
Traditional automation executes predefined sequences, such as setting a target temperature, opening a source shutter at a specified time, and growing a layer for a prescribed duration.
Autonomous epitaxy introduces an additional decision-making layer. The system evaluates the current process state and uses feedback to inform subsequent actions.
Recent research has explored self-driving MBE platforms that integrate in-situ RHEED, equipment control, and physics-informed machine learning. Earlier work has also demonstrated machine-learning-assisted feedback control for InAs/GaAs quantum-dot growth using RHEED video.
These demonstrations establish important technical directions, but they should not be confused with proof that fully autonomous, high-volume GaN manufacturing has already been achieved across all production conditions.
Industrial deployment requires reliable sensors, representative datasets, validated models, safe control logic, and consistent performance across equipment operating cycles.
For buyers, the use of AI in manufacturing can be informative, but the final wafer specification remains the primary basis for qualification.
A GaN epi-wafer request for quotation should clearly define the following requirements.
Wafer size: 2-inch, 4-inch, 6-inch, or 8-inch formats, depending on supplier capability and application.
Substrate platform: GaN-on-Si, GaN-on-SiC, GaN-on-sapphire, or free-standing GaN.
Epitaxial structure: GaN templates, AlGaN/GaN high-electron-mobility transistor structures, power-device epitaxy, LED structures, or customized research layers.
Layer thickness: Target thicknesses and tolerances for each relevant layer.
Electrical properties: Sheet resistance, carrier concentration, mobility, buffer leakage, and breakdown-related requirements where applicable.
Surface quality: RMS roughness, particle limits, surface pits, and scratch criteria.
Wafer geometry: Total thickness variation (TTV), bow, and warp.
Uniformity: Center-to-edge variation and wafer-to-wafer consistency should be defined separately from average values.
Post-growth characterization remains essential, even when the growth process uses advanced AI monitoring. Surface inspection, AFM, XRD, electrical measurements, thickness mapping, and wafer-geometry measurements provide independent evidence that the delivered material meets its specification.
AI-controlled GaN epitaxy represents a meaningful evolution in semiconductor process engineering. By combining RHEED image analysis, machine learning, and equipment sensor data, manufacturers can move toward more quantitative monitoring, earlier detection of process drift, and increasingly adaptive MBE control.
The potential benefits include improved run-to-run consistency, more stable growth conditions, better process-window management, and opportunities to improve wafer uniformity and effective yield.
However, AI is not a substitute for sound reactor design, high-quality substrates, appropriate buffer-layer engineering, accurate calibration, or rigorous materials characterization.
For GaN epi-wafer manufacturers, the real opportunity lies in integrating data-driven intelligence with established materials science. For buyers, the ultimate test remains measurable and reproducible wafer quality—not the presence of AI alone.
Gallium nitride (GaN) epitaxy is entering a new phase of semiconductor manufacturing. As demand grows for high-performance RF devices and power electronics, manufacturers are placing greater emphasis on epitaxial quality, process stability, and wafer-to-wafer consistency.
Traditionally, GaN epitaxy has relied on carefully developed growth recipes, experienced process engineers, and continuous monitoring of reactor conditions. However, epitaxial growth involves numerous interacting variables. Even minor changes in substrate temperature, chamber pressure, material flux, or surface reconstruction can influence the resulting crystal structure and electrical properties.
Artificial intelligence (AI) is opening up a new approach. By combining in-situ Reflection High-Energy Electron Diffraction (RHEED), equipment sensor data, and machine-learning algorithms, manufacturers can move beyond fixed growth recipes toward more adaptive, data-driven process control.
In September 2026, South Korea's IVWorks reported an AI-enabled GaN epitaxy manufacturing system that analyzes RHEED data alongside molecular beam epitaxy (MBE) equipment information. According to the company, its accumulated datasets include more than 330 million time-series process data points and over 30 TB of RHEED data collected since 2019.
The development illustrates a broader industry trend: semiconductor epitaxy is gradually moving from recipe-based automation toward increasingly intelligent and potentially autonomous manufacturing.
GaN is widely used in RF communications, high-frequency electronics, and power semiconductor applications. In these devices, epitaxial quality directly affects electrical performance, reliability, and manufacturing yield.
Important GaN epi-wafer characteristics include:
Epitaxial layer thickness and uniformity
Surface roughness and morphology
Crystal defect density
Interface quality
Carrier concentration and electron mobility
Sheet resistance and its uniformity
Buffer leakage characteristics
Wafer stress, bow, and warp
Particle contamination
Achieving the required properties involves controlling several interacting growth parameters, including substrate temperature, gallium and nitrogen fluxes, growth rate, chamber pressure, plasma conditions, doping flux, growth time, and surface reconstruction.
These parameters cannot always be optimized independently. A temperature change, for example, may affect surface diffusion, material incorporation, and growth kinetics simultaneously. The resulting variation may not be immediately visible at the equipment level, yet it can influence the final wafer's electrical or structural characteristics.
This complexity makes real-time process monitoring and early detection of abnormal growth conditions particularly valuable.
Reflection High-Energy Electron Diffraction is an established in-situ characterization technique used in MBE.
During growth, a high-energy electron beam strikes the wafer surface at a shallow angle. The resulting diffraction pattern provides information about the crystal surface without requiring the wafer to be removed from the reactor.
Depending on the growth conditions, RHEED patterns can reveal changes associated with:
Surface reconstruction and crystalline order
Growth mode and surface morphology
Surface roughening
Layer formation and growth transitions
For example, relatively smooth two-dimensional growth can produce characteristic diffraction streaks. As the surface becomes rougher or three-dimensional islands develop, the pattern may change.
RHEED is therefore useful not only for observing the growth process but also for identifying changes that may indicate a transition away from the intended growth regime.
However, interpreting a continuous stream of diffraction images presents a challenge. Conventional visual inspection depends heavily on operator experience, and subtle changes may be difficult to quantify consistently across long growth runs.
Machine learning offers a way to analyze RHEED data systematically and at scale.
Instead of relying exclusively on human interpretation of individual images, an AI model can evaluate image sequences and identify changes in diffraction intensity, streak width, spot formation, pattern geometry, and temporal evolution.
A study published in the Journal of Vacuum Science & Technology A demonstrated machine-learning analysis of RHEED images during MBE growth to identify changes in deposition regimes in real time. Its analysis workflow was designed to operate in less than one second, illustrating the potential for rapid diffraction analysis within a process-monitoring system.
The key advantage is the ability to convert complex image sequences into measurable indicators of growth behavior.
When these indicators are combined with reactor sensor data, the system can examine relationships between observed surface conditions and operating parameters. This creates opportunities to detect abnormal growth states earlier and compare current growth behavior with historical process data.
Nevertheless, AI predictions depend on the quality and representativeness of the training data. A model developed for one material system or reactor configuration cannot automatically be assumed to perform equally well in another.
AI-assisted RHEED analysis becomes particularly valuable when it is integrated into a feedback-control architecture.
A simplified closed-loop system follows this sequence:
A camera captures RHEED diffraction patterns during growth.
Image-processing software extracts relevant features.
A machine-learning model evaluates the current growth state.
The control system determines whether the process is operating within the intended window.
Where validated control rules permit, selected MBE parameters are adjusted.
Subsequent RHEED data are analyzed to evaluate the response.
This approach differs from conventional automation, which primarily executes predefined instructions. A feedback-enabled system can use current process information to inform subsequent decisions.
For example, a model may identify diffraction changes associated with a transition from smooth layer-by-layer growth toward a less desirable growth mode. A supervisory controller could then assess whether an adjustment to substrate temperature or material flux is appropriate.
Such adjustments must remain within a validated process window. AI-based monitoring does not eliminate the need for physical process models, engineering constraints, or safeguards against unstable control actions.
The longer-term objective is increasingly autonomous epitaxy, in which monitoring, decision-making, and process adjustment are integrated into a coordinated manufacturing system.
Uniformity is a critical requirement for commercial GaN epi wafers, particularly as manufacturers move toward larger wafer formats.
Depending on the device structure, customers may evaluate thickness uniformity, sheet resistance, carrier concentration, mobility, aluminum composition, surface morphology, and wafer bow.
AI-assisted process control may contribute to better uniformity in several ways.
MBE reactor conditions can change over time because of source depletion, chamber coating, equipment aging, temperature-calibration changes, and plasma-source variation.
By analyzing RHEED patterns alongside temperature, pressure, and other equipment data, an AI system may identify changes in process behavior before they become obvious in conventional monitoring.
Changes in diffraction patterns can provide early indications that the surface is entering an undesirable growth regime. Timely detection allows engineers to investigate the cause and, where appropriate, intervene before the deviation becomes more severe.
Historical process records can be used to compare current growth trajectories with previously characterized runs. This may help identify deviations associated with changes in electrical properties, surface morphology, or other wafer-quality indicators.
These benefits remain dependent on validation. Improved process monitoring does not automatically guarantee better thickness uniformity, lower defect density, or higher production yield; those outcomes must be confirmed through post-growth measurements.
An effective AI process-control platform requires more than a large collection of sensor readings.
Useful datasets may include RHEED videos, substrate temperature, chamber pressure, material flux, plasma power, source temperature, growth rate, growth time, and maintenance history.
These records become more valuable when linked to post-growth characterization, including:
X-ray diffraction (XRD)
Atomic force microscopy (AFM)
Hall-effect measurements
Sheet-resistance mapping
Photoluminescence
Surface inspection
Epitaxial thickness mapping
The objective is to establish a reliable relationship between process conditions, in-situ surface behavior, and final wafer properties.
For example, if a particular combination of diffraction-pattern evolution and reactor conditions repeatedly correlates with an unfavorable post-growth result, the model may learn to flag similar conditions in subsequent runs.
However, correlation alone does not prove causation. Process engineers must verify whether the identified relationships are physically meaningful and reproducible before using them to guide production decisions.
GaN epitaxial defects can include threading dislocations, surface pits, hillocks, cracks, particles, interface defects, local roughness, and non-uniform doping.
Many defects are influenced by substrate quality, lattice mismatch, thermal stress, buffer-layer design, and other fundamental materials considerations. AI cannot eliminate these physical limitations.
Its more immediate opportunity is to help control defects associated with process instability.
By identifying recurring patterns in in-situ measurements and comparing them with post-growth inspection results, an AI platform may help manufacturers recognize process conditions associated with elevated defect risk.
This can support earlier intervention, more targeted troubleshooting, and improved process-window management.
The practical objective is not to replace materials science, but to maintain growth conditions within a narrower and more repeatable operating range.
Traditional automation executes predefined sequences, such as setting a target temperature, opening a source shutter at a specified time, and growing a layer for a prescribed duration.
Autonomous epitaxy introduces an additional decision-making layer. The system evaluates the current process state and uses feedback to inform subsequent actions.
Recent research has explored self-driving MBE platforms that integrate in-situ RHEED, equipment control, and physics-informed machine learning. Earlier work has also demonstrated machine-learning-assisted feedback control for InAs/GaAs quantum-dot growth using RHEED video.
These demonstrations establish important technical directions, but they should not be confused with proof that fully autonomous, high-volume GaN manufacturing has already been achieved across all production conditions.
Industrial deployment requires reliable sensors, representative datasets, validated models, safe control logic, and consistent performance across equipment operating cycles.
For buyers, the use of AI in manufacturing can be informative, but the final wafer specification remains the primary basis for qualification.
A GaN epi-wafer request for quotation should clearly define the following requirements.
Wafer size: 2-inch, 4-inch, 6-inch, or 8-inch formats, depending on supplier capability and application.
Substrate platform: GaN-on-Si, GaN-on-SiC, GaN-on-sapphire, or free-standing GaN.
Epitaxial structure: GaN templates, AlGaN/GaN high-electron-mobility transistor structures, power-device epitaxy, LED structures, or customized research layers.
Layer thickness: Target thicknesses and tolerances for each relevant layer.
Electrical properties: Sheet resistance, carrier concentration, mobility, buffer leakage, and breakdown-related requirements where applicable.
Surface quality: RMS roughness, particle limits, surface pits, and scratch criteria.
Wafer geometry: Total thickness variation (TTV), bow, and warp.
Uniformity: Center-to-edge variation and wafer-to-wafer consistency should be defined separately from average values.
Post-growth characterization remains essential, even when the growth process uses advanced AI monitoring. Surface inspection, AFM, XRD, electrical measurements, thickness mapping, and wafer-geometry measurements provide independent evidence that the delivered material meets its specification.
AI-controlled GaN epitaxy represents a meaningful evolution in semiconductor process engineering. By combining RHEED image analysis, machine learning, and equipment sensor data, manufacturers can move toward more quantitative monitoring, earlier detection of process drift, and increasingly adaptive MBE control.
The potential benefits include improved run-to-run consistency, more stable growth conditions, better process-window management, and opportunities to improve wafer uniformity and effective yield.
However, AI is not a substitute for sound reactor design, high-quality substrates, appropriate buffer-layer engineering, accurate calibration, or rigorous materials characterization.
For GaN epi-wafer manufacturers, the real opportunity lies in integrating data-driven intelligence with established materials science. For buyers, the ultimate test remains measurable and reproducible wafer quality—not the presence of AI alone.
english
français
Deutsch
Italiano
Русский
Español
português
Nederlandse
ελληνικά
日本語
한국
العربية
हिन्दी
Türkçe
indonesia
tiếng Việt
ไทย
বাংলা
فارسی
polski