Gallium nitride epitaxy is entering a new stage of semiconductor manufacturing.
For years, GaN epi-wafer quality has depended heavily on carefully optimized growth recipes, experienced process engineers and continuous monitoring of reactor conditions. Even small deviations in substrate temperature, chamber pressure, III/V flux ratio or surface reconstruction can influence the final crystal quality.
Artificial intelligence is beginning to change this approach.
Instead of relying only on fixed recipes, new epitaxy platforms are combining real-time RHEED data, equipment sensor information and machine-learning algorithms to identify growth conditions, detect abnormalities and ultimately adjust process parameters during epitaxial growth.
A recent example comes from South Korea's IVWorks, which reported in September 2026 that its GaN epi-wafer production system uses AI models to analyze Reflection High-Energy Electron Diffraction, or RHEED, together with MBE equipment data. The company says it has accumulated more than 330 million time-series process data points and over 30 TB of RHEED data since 2019.
This development highlights a broader trend: GaN epitaxy is moving from recipe-based automation toward data-driven and increasingly autonomous manufacturing.
GaN devices are used in applications ranging from RF communications to high-frequency power electronics.
However, the electrical performance of the finished device depends strongly on epitaxial quality.
Important GaN epi-wafer characteristics include:
These parameters are influenced by many interacting growth conditions.
During Molecular Beam Epitaxy, for example, engineers must control variables such as:
The difficulty is that these variables are not independent.
Changing temperature may affect surface mobility, incorporation efficiency and growth rate simultaneously.
A small deviation that appears insignificant at the equipment level can eventually produce measurable differences in wafer electrical performance.
This is one reason AI-based process monitoring is attracting attention.
Reflection High-Energy Electron Diffraction is one of the most useful in-situ monitoring techniques for MBE.
During epitaxial growth, a high-energy electron beam strikes the wafer surface at a very shallow angle.
The electrons interact with the crystal surface and create a diffraction pattern.
Engineers can use this pattern to obtain information about the surface during growth without removing the wafer from the reactor.
RHEED can provide information related to:
Different surface conditions produce different diffraction patterns.
A smooth two-dimensional surface often produces relatively sharp streaks, while roughening or three-dimensional island formation can change the diffraction pattern significantly.
Traditionally, experienced operators interpret these patterns visually.
That approach works, but it has limitations.
A modern epitaxy run can generate an enormous amount of RHEED information.
A camera may continuously record diffraction patterns throughout the growth process.
Human operators cannot realistically analyze every frame quantitatively.
Some changes may also occur before they become visually obvious.
Machine learning offers a different approach.
Instead of asking an engineer to interpret individual images, AI systems can analyze entire RHEED image sequences and identify subtle changes in:
Research published in the Journal of Vacuum Science & Technology A demonstrated machine-learning analysis of RHEED images during MBE growth that could identify changes in deposition regimes on the fly. The analysis workflow was designed to operate in less than one second, illustrating the potential for RHEED data to become part of real-time feedback control rather than simply a monitoring tool.
The important change is not simply using AI to classify images.
The larger goal is closed-loop process control.
A simplified autonomous epitaxy system can operate as follows:
RHEED camera
↓
Real-time image analysis
↓
Machine-learning model
↓
Growth-state classification
↓
Process decision
↓
MBE parameter adjustment
↓
New RHEED response
The cycle then repeats.
This allows the growth system to respond to the actual condition of the wafer rather than blindly following a predetermined recipe.
For example, a model could potentially detect that a surface is beginning to transition from smooth layer-by-layer growth toward undesirable island formation.
The process controller could then evaluate whether parameters such as substrate temperature or material flux should be adjusted.
This concept is increasingly described as autonomous epitaxy or self-driving epitaxy.
The September 2026 IVWorks announcement provides an industrial GaN example.
According to the company, its AI epitaxy platform combines RHEED information with time-series equipment data such as temperature and pressure.
AI models evaluate the growth condition and detect abnormalities, while its MBE automation framework handles operations including wafer loading, epitaxial growth and unloading. The stated direction is toward reducing human intervention and improving production stability.
This is important because industrial epitaxy is different from laboratory growth.
Mass production requires repeatability across:
An excellent single wafer is not enough.
A commercial epi-wafer supplier needs consistent performance wafer after wafer.
Uniformity is one of the most important specifications for commercial GaN epi wafers.
Depending on the device structure, customers may evaluate:
For a small research sample, local variations may sometimes be acceptable.
For 100mm, 150mm or 200mm production wafers, they can directly affect device yield.
AI process control can potentially improve uniformity in several ways.
MBE reactors do not remain perfectly unchanged.
Conditions can drift because of:
A fixed recipe may therefore produce slightly different results after many growth cycles.
By combining RHEED data with equipment sensor data, AI systems may detect drift earlier than conventional statistical monitoring.
An abnormal RHEED pattern may indicate that the surface is entering an undesirable growth regime.
Early detection allows the system to flag the wafer or potentially adjust the process before the defect becomes severe.
A major manufacturing challenge is maintaining the same electrical characteristics between batches.
Machine-learning models can compare the current run with large historical datasets and identify deviations from previously successful growth trajectories.
This could help improve run-to-run repeatability.
Surface morphology strongly affects GaN device fabrication.
Poor surface quality can influence:
RHEED is especially valuable because it monitors the crystal surface directly during growth.
Machine-learning-assisted RHEED analysis has already demonstrated that diffraction information can be correlated with important film-quality indicators in experimental MBE systems.
For example, a 2025 study of GaSe MBE growth showed that incorporating RHEED-derived features improved machine-learning models used to predict crystal-quality metrics, supporting the broader concept of combining in-situ diffraction with process data for epitaxy optimization.
The specific material in that study was not GaN, so the quantitative model cannot simply be transferred to GaN production. However, the approach demonstrates how RHEED information can become a machine-readable process variable rather than only a visual diagnostic.
Yield is ultimately one of the most important reasons to introduce advanced process control.
GaN epitaxial defects may include:
Not every defect can be eliminated through AI.
Many originate from fundamental substrate and heteroepitaxy limitations.
However, AI may help reduce defects related to process instability.
For example, if a particular combination of RHEED pattern evolution, temperature drift and chamber pressure repeatedly correlates with poor post-growth results, the model can learn that relationship.
Future production runs could then flag similar conditions before the wafer is completed.
The objective is not for AI to change the laws of crystal growth.
It is to keep the growth process inside a narrower and more repeatable process window.
AI-controlled epitaxy depends heavily on data quality.
A model trained on only a small number of growth runs may not understand enough process conditions to make reliable decisions.
This explains why large historical datasets are valuable.
Useful information may include:
Post-growth characterization can then be linked with these process records.
Examples include:
This creates a relationship between:
Process conditions → in-situ surface behavior → final wafer quality
Once enough reliable data is available, the system can begin finding relationships that may be difficult to identify manually.
Traditional semiconductor automation follows predefined instructions.
For example:
Set temperature to X.
Open source shutter at time Y.
Grow layer for Z minutes.
Autonomous manufacturing adds another level.
Instead of only executing instructions, the system evaluates the current process condition before deciding what should happen next.
Recent research has already demonstrated self-driving MBE concepts using in-situ RHEED and physics-informed machine learning. A 2026 study described the SemiEpi platform, which integrates MBE equipment, RHEED monitoring and feedback control to adjust semiconductor heterostructure growth.
Earlier research has also demonstrated machine-learning-assisted real-time control of InAs/GaAs quantum-dot growth using RHEED video as the input for feedback decisions.
These demonstrations are not equivalent to fully autonomous high-volume GaN manufacturing, but they show the technical path toward increasingly closed-loop epitaxy.
AI does not remove the need for materials science.
High-quality GaN epitaxy still depends on:
An AI model trained on poor process data will not automatically produce excellent GaN crystals.
The physical growth process still determines the final material.
AI mainly provides a more powerful way to monitor, analyze and control that process.
For customers purchasing GaN epi wafers, AI-controlled manufacturing itself should not be the primary acceptance criterion.
The final wafer specification remains more important.
A GaN epi-wafer RFQ should clearly define relevant parameters.
Common options may include:
Possible platforms include:
Examples include:
Specify required layer thickness and tolerance whenever possible.
Depending on the application, specifications may include:
Customers may specify:
For larger wafers, important parameters include:
For device manufacturing, center-to-edge and wafer-to-wafer uniformity should be discussed separately from average values.
The semiconductor industry is steadily moving GaN production toward larger wafer formats.
As wafer diameter increases, a small process gradient can affect a much larger number of devices.
For example, variations in:
may produce radial differences across the wafer.
This is why data-driven process control becomes increasingly valuable for 150mm and 200mm GaN epitaxy and potentially future larger-diameter production.
The economic value of AI is therefore closely related to wafer scale.
If improved process control reduces wafer-to-wafer variation or increases the percentage of usable wafer area, the impact can be multiplied across thousands of devices.
Even a highly automated growth process still requires post-growth characterization.
Critical GaN epi wafers may require a combination of:
AI-controlled growth and post-growth metrology should therefore be considered complementary.
The growth system provides real-time information.
Final inspection verifies whether the wafer meets the customer's specification.
The long-term direction of semiconductor epitaxy may go beyond simple anomaly detection.
Future platforms could combine:
A digital representation of the epitaxy process could continuously compare predicted growth behavior with actual reactor data.
The system could then optimize process parameters for specific targets such as:
This could be especially valuable for custom GaN epitaxy, where customers require different device structures and electrical specifications.
AI-controlled epitaxy is changing how GaN epi-wafer manufacturing can be monitored and optimized.
RHEED has traditionally served as an important real-time observation tool for MBE growth. By combining RHEED images with machine learning and equipment sensor data, manufacturers can transform this information into a more quantitative process-control system.
The potential benefits include:
Recent industrial developments in GaN manufacturing and academic demonstrations of self-driving MBE indicate that semiconductor epitaxy is gradually moving from fixed recipe automation toward increasingly autonomous process control.
For GaN epi-wafer customers, however, the most important factors remain measurable material specifications: epitaxial structure, electrical properties, surface quality, wafer geometry and wafer-to-wafer uniformity.
AI is not a substitute for high-quality GaN epitaxy.
Its real value is helping manufacturers achieve that quality more consistently, across more wafers and with tighter control over increasingly complex semiconductor growth processes.
Gallium nitride epitaxy is entering a new stage of semiconductor manufacturing.
For years, GaN epi-wafer quality has depended heavily on carefully optimized growth recipes, experienced process engineers and continuous monitoring of reactor conditions. Even small deviations in substrate temperature, chamber pressure, III/V flux ratio or surface reconstruction can influence the final crystal quality.
Artificial intelligence is beginning to change this approach.
Instead of relying only on fixed recipes, new epitaxy platforms are combining real-time RHEED data, equipment sensor information and machine-learning algorithms to identify growth conditions, detect abnormalities and ultimately adjust process parameters during epitaxial growth.
A recent example comes from South Korea's IVWorks, which reported in September 2026 that its GaN epi-wafer production system uses AI models to analyze Reflection High-Energy Electron Diffraction, or RHEED, together with MBE equipment data. The company says it has accumulated more than 330 million time-series process data points and over 30 TB of RHEED data since 2019.
This development highlights a broader trend: GaN epitaxy is moving from recipe-based automation toward data-driven and increasingly autonomous manufacturing.
GaN devices are used in applications ranging from RF communications to high-frequency power electronics.
However, the electrical performance of the finished device depends strongly on epitaxial quality.
Important GaN epi-wafer characteristics include:
These parameters are influenced by many interacting growth conditions.
During Molecular Beam Epitaxy, for example, engineers must control variables such as:
The difficulty is that these variables are not independent.
Changing temperature may affect surface mobility, incorporation efficiency and growth rate simultaneously.
A small deviation that appears insignificant at the equipment level can eventually produce measurable differences in wafer electrical performance.
This is one reason AI-based process monitoring is attracting attention.
Reflection High-Energy Electron Diffraction is one of the most useful in-situ monitoring techniques for MBE.
During epitaxial growth, a high-energy electron beam strikes the wafer surface at a very shallow angle.
The electrons interact with the crystal surface and create a diffraction pattern.
Engineers can use this pattern to obtain information about the surface during growth without removing the wafer from the reactor.
RHEED can provide information related to:
Different surface conditions produce different diffraction patterns.
A smooth two-dimensional surface often produces relatively sharp streaks, while roughening or three-dimensional island formation can change the diffraction pattern significantly.
Traditionally, experienced operators interpret these patterns visually.
That approach works, but it has limitations.
A modern epitaxy run can generate an enormous amount of RHEED information.
A camera may continuously record diffraction patterns throughout the growth process.
Human operators cannot realistically analyze every frame quantitatively.
Some changes may also occur before they become visually obvious.
Machine learning offers a different approach.
Instead of asking an engineer to interpret individual images, AI systems can analyze entire RHEED image sequences and identify subtle changes in:
Research published in the Journal of Vacuum Science & Technology A demonstrated machine-learning analysis of RHEED images during MBE growth that could identify changes in deposition regimes on the fly. The analysis workflow was designed to operate in less than one second, illustrating the potential for RHEED data to become part of real-time feedback control rather than simply a monitoring tool.
The important change is not simply using AI to classify images.
The larger goal is closed-loop process control.
A simplified autonomous epitaxy system can operate as follows:
RHEED camera
↓
Real-time image analysis
↓
Machine-learning model
↓
Growth-state classification
↓
Process decision
↓
MBE parameter adjustment
↓
New RHEED response
The cycle then repeats.
This allows the growth system to respond to the actual condition of the wafer rather than blindly following a predetermined recipe.
For example, a model could potentially detect that a surface is beginning to transition from smooth layer-by-layer growth toward undesirable island formation.
The process controller could then evaluate whether parameters such as substrate temperature or material flux should be adjusted.
This concept is increasingly described as autonomous epitaxy or self-driving epitaxy.
The September 2026 IVWorks announcement provides an industrial GaN example.
According to the company, its AI epitaxy platform combines RHEED information with time-series equipment data such as temperature and pressure.
AI models evaluate the growth condition and detect abnormalities, while its MBE automation framework handles operations including wafer loading, epitaxial growth and unloading. The stated direction is toward reducing human intervention and improving production stability.
This is important because industrial epitaxy is different from laboratory growth.
Mass production requires repeatability across:
An excellent single wafer is not enough.
A commercial epi-wafer supplier needs consistent performance wafer after wafer.
Uniformity is one of the most important specifications for commercial GaN epi wafers.
Depending on the device structure, customers may evaluate:
For a small research sample, local variations may sometimes be acceptable.
For 100mm, 150mm or 200mm production wafers, they can directly affect device yield.
AI process control can potentially improve uniformity in several ways.
MBE reactors do not remain perfectly unchanged.
Conditions can drift because of:
A fixed recipe may therefore produce slightly different results after many growth cycles.
By combining RHEED data with equipment sensor data, AI systems may detect drift earlier than conventional statistical monitoring.
An abnormal RHEED pattern may indicate that the surface is entering an undesirable growth regime.
Early detection allows the system to flag the wafer or potentially adjust the process before the defect becomes severe.
A major manufacturing challenge is maintaining the same electrical characteristics between batches.
Machine-learning models can compare the current run with large historical datasets and identify deviations from previously successful growth trajectories.
This could help improve run-to-run repeatability.
Surface morphology strongly affects GaN device fabrication.
Poor surface quality can influence:
RHEED is especially valuable because it monitors the crystal surface directly during growth.
Machine-learning-assisted RHEED analysis has already demonstrated that diffraction information can be correlated with important film-quality indicators in experimental MBE systems.
For example, a 2025 study of GaSe MBE growth showed that incorporating RHEED-derived features improved machine-learning models used to predict crystal-quality metrics, supporting the broader concept of combining in-situ diffraction with process data for epitaxy optimization.
The specific material in that study was not GaN, so the quantitative model cannot simply be transferred to GaN production. However, the approach demonstrates how RHEED information can become a machine-readable process variable rather than only a visual diagnostic.
Yield is ultimately one of the most important reasons to introduce advanced process control.
GaN epitaxial defects may include:
Not every defect can be eliminated through AI.
Many originate from fundamental substrate and heteroepitaxy limitations.
However, AI may help reduce defects related to process instability.
For example, if a particular combination of RHEED pattern evolution, temperature drift and chamber pressure repeatedly correlates with poor post-growth results, the model can learn that relationship.
Future production runs could then flag similar conditions before the wafer is completed.
The objective is not for AI to change the laws of crystal growth.
It is to keep the growth process inside a narrower and more repeatable process window.
AI-controlled epitaxy depends heavily on data quality.
A model trained on only a small number of growth runs may not understand enough process conditions to make reliable decisions.
This explains why large historical datasets are valuable.
Useful information may include:
Post-growth characterization can then be linked with these process records.
Examples include:
This creates a relationship between:
Process conditions → in-situ surface behavior → final wafer quality
Once enough reliable data is available, the system can begin finding relationships that may be difficult to identify manually.
Traditional semiconductor automation follows predefined instructions.
For example:
Set temperature to X.
Open source shutter at time Y.
Grow layer for Z minutes.
Autonomous manufacturing adds another level.
Instead of only executing instructions, the system evaluates the current process condition before deciding what should happen next.
Recent research has already demonstrated self-driving MBE concepts using in-situ RHEED and physics-informed machine learning. A 2026 study described the SemiEpi platform, which integrates MBE equipment, RHEED monitoring and feedback control to adjust semiconductor heterostructure growth.
Earlier research has also demonstrated machine-learning-assisted real-time control of InAs/GaAs quantum-dot growth using RHEED video as the input for feedback decisions.
These demonstrations are not equivalent to fully autonomous high-volume GaN manufacturing, but they show the technical path toward increasingly closed-loop epitaxy.
AI does not remove the need for materials science.
High-quality GaN epitaxy still depends on:
An AI model trained on poor process data will not automatically produce excellent GaN crystals.
The physical growth process still determines the final material.
AI mainly provides a more powerful way to monitor, analyze and control that process.
For customers purchasing GaN epi wafers, AI-controlled manufacturing itself should not be the primary acceptance criterion.
The final wafer specification remains more important.
A GaN epi-wafer RFQ should clearly define relevant parameters.
Common options may include:
Possible platforms include:
Examples include:
Specify required layer thickness and tolerance whenever possible.
Depending on the application, specifications may include:
Customers may specify:
For larger wafers, important parameters include:
For device manufacturing, center-to-edge and wafer-to-wafer uniformity should be discussed separately from average values.
The semiconductor industry is steadily moving GaN production toward larger wafer formats.
As wafer diameter increases, a small process gradient can affect a much larger number of devices.
For example, variations in:
may produce radial differences across the wafer.
This is why data-driven process control becomes increasingly valuable for 150mm and 200mm GaN epitaxy and potentially future larger-diameter production.
The economic value of AI is therefore closely related to wafer scale.
If improved process control reduces wafer-to-wafer variation or increases the percentage of usable wafer area, the impact can be multiplied across thousands of devices.
Even a highly automated growth process still requires post-growth characterization.
Critical GaN epi wafers may require a combination of:
AI-controlled growth and post-growth metrology should therefore be considered complementary.
The growth system provides real-time information.
Final inspection verifies whether the wafer meets the customer's specification.
The long-term direction of semiconductor epitaxy may go beyond simple anomaly detection.
Future platforms could combine:
A digital representation of the epitaxy process could continuously compare predicted growth behavior with actual reactor data.
The system could then optimize process parameters for specific targets such as:
This could be especially valuable for custom GaN epitaxy, where customers require different device structures and electrical specifications.
AI-controlled epitaxy is changing how GaN epi-wafer manufacturing can be monitored and optimized.
RHEED has traditionally served as an important real-time observation tool for MBE growth. By combining RHEED images with machine learning and equipment sensor data, manufacturers can transform this information into a more quantitative process-control system.
The potential benefits include:
Recent industrial developments in GaN manufacturing and academic demonstrations of self-driving MBE indicate that semiconductor epitaxy is gradually moving from fixed recipe automation toward increasingly autonomous process control.
For GaN epi-wafer customers, however, the most important factors remain measurable material specifications: epitaxial structure, electrical properties, surface quality, wafer geometry and wafer-to-wafer uniformity.
AI is not a substitute for high-quality GaN epitaxy.
Its real value is helping manufacturers achieve that quality more consistently, across more wafers and with tighter control over increasingly complex semiconductor growth processes.
english
français
Deutsch
Italiano
Русский
Español
português
Nederlandse
ελληνικά
日本語
한국
العربية
Türkçe
polski