Key Takeaways:
AI can identify promising process conditions and recognize manufacturing problems, but those results still require physical validation before they can be trusted in production.
Incomplete manufacturing data and uncertainty about downstream effects make it difficult to give AI systems authority over process adjustments.
Greater manufacturing autonomy will require clear decision boundaries, failure containment, and ways to qualify models that change as they learn.
Semiconductor manufacturing has always depended on engineers making decisions involving complex interactions among process parameters. An etch recipe, for example, may involve combinations of RF power, pressure, gas chemistry, temperature, and timing, each affecting multiple process characteristics. Changing one parameter might improve critical dimension control while creating problems elsewhere. But as devices and manufacturing processes become more complicated, the growing number of variables involved makes those decisions increasingly difficult.
AI-driven manufacturing offers a way to manage some of that complexity, from finding better process conditions to identifying defects and responding to problems on the production line. But how much production responsibility engineers are willing to give those systems remains an open question, and one that drew considerable attention at SEMICON West 2026.
“No matter how good we are as engineers, no matter how disciplined and objective we are, we tend to approach these difficult problems with historical biases,” said David Fried, chief AI officer and corporate vice president of Semiverse Solutions at Lam Research during a conference panel. “A process engineer might look at a set of data and say, ‘I need to move into a lower-pressure, higher-power regime because that’s how we solved this problem 10 years ago.’ But with this massive parameter space, AI can point us in a different direction, into an operating region we might never have considered because we couldn’t see those relationships in the data before.”
Finding a process that keeps working
Finding an unfamiliar operating condition is useful, but determining whether it will hold up in production is another matter. Two recipes may produce similar results during development, yet respond quite differently to changes in incoming materials or equipment conditions. AI can help engineers identify those differences before committing to a particular process.
“It’s not just a question of delivering to a particular performance point. It’s also a question of how stable that particular performance point is, how robust it is,” said Vikas Sharma, senior principal engineer and senior director of pathfinding for Intel Foundry R&D, in a recent presentation. “Minor perturbations in the incoming material or at the tool can lead to a significant drop in performance.”
The scale of the optimization problem helps explain why AI-driven manufacturing is attracting so much interest. Fried described a current Lam Research etcher with a potential recipe space on the order of 10²³ combinations. At Intel, Sharma discussed a process involving 350 parameters and approximately 10⁷⁰⁰ theoretical combinations.
According to Sharma, engineers working with equipment suppliers spent about a year developing a minimally acceptable recipe for that particular tool and process layer, followed by an extended effort to improve it with limited success. After introducing AI-assisted optimization, they began finding better results within a few months, eventually identifying multiple solutions that improved on the original baseline.
The challenge was determining which solutions were worth pursuing. Sharma noted that AI could identify thousands of candidates, but evaluating them still required expensive silicon experiments. Engineers also had to account for nonlinear relationships among parameters, which could change considerably across different operating regions.
“AI is not a magic wand. You can’t use the same AI solution for every problem,” Fried said. “Figuring out whether you’re using full physical simulation, surrogate models, machine learning, deep learning, LLM-based AI, or agentic AI, which of these solutions you’re building in which form to solve which problem and address which workflow, that’s more of what we’re spending our time on than actually building it or implementing it.”
The different approaches serve different purposes, from exploring process conditions to coordinating engineering workflows. Combining them may allow engineers to investigate more possibilities, but it also raises questions about how the results should be validated and whether the assumptions used in one model remain appropriate when its recommendations are passed to another.
When the factory acts on what it sees
Inspection is one of the more established applications of AI in manufacturing, particularly in advanced packaging, where finer pitches and more complicated structures have increased the number of connections that must be examined. Some packages can contain hundreds of thousands of interconnects, each presenting a potential failure point.
Automated optical inspection and X-ray systems collect extensive information about those structures. Machine learning can identify variations that might be difficult to distinguish visually, but manufacturers still need to determine which differences indicate actual defects.
“When we look at X-rays, everybody just sees all shades of gray, but AI can see a thousand shades of gray,” said Ingu Yin Chang, executive vice president at ASE Group. “But the AI only knows if this shade of gray equates to a problem once we say, ‘Hey, this equates to a problem.’ So, you’ve got to have the physical data that says, ‘Okay, that is a problem.’ Then, in the future, they can scan for that shade of gray and that location.”
Once manufacturers determine which image characteristics correspond to actual defects, AI-enabled inspection systems can flag abnormalities during production and trigger predefined equipment responses. At ASE, systems can recognize known abnormalities and trigger equipment responses before additional assemblies are affected.
“We have figured out how to do a lot of automation and use AI to help us identify potential issues, based on the issues that we physically learned through traditional manufacturing methods,” Chang said. “We will know when a wire bond is not bonded correctly, and the machine will immediately stop. There’s constant AI oversight to ask, ‘Are we doing everything correctly? Are we seeing it as we planned to see it?’”
Stopping equipment when a defect is detected can prevent additional losses, but it doesn’t necessarily mean AI is controlling the process independently. The inspection system may identify an abnormal condition while a conventional controller executes a predefined response. Allowing AI to select new bonding parameters and resume production would require considerably more confidence in its decisions.
ASE’s product diversity makes that challenge more difficult. Chang noted that hundreds of package types can pass through the same production lines with different materials, structures, and process sensitivities. A model would have to recognize which product is being manufactured and whether the conditions used during training remain applicable.
An inspection system might correctly identify a defective bond, but changing the process to prevent another failure requires knowing what caused it and whether an adjustment will create problems elsewhere. Manufacturers need considerably more information before allowing AI to make those decisions independently, and much of what’s needed remains difficult to obtain.
Proving the model is right
One of the first obstacles to AI-driven manufacturing is establishing whether an AI recommendation accurately represents the physical process. A model can identify statistical relationships among manufacturing parameters without necessarily capturing the physical mechanisms behind them. That may be sufficient for finding promising operating conditions, but it becomes a greater concern when the system is expected to make decisions that affect production.
Sharma’s Intel example illustrates the problem. Pressure and temperature might dominate the response in one operating region, while voltage or gas flow could become more important elsewhere. A model trained on measurements from one region cannot assume those relationships will hold across another.
“The fundamental principle is that surrogate models are reliable when interpolating within the region represented by their training data and less reliable when extrapolating beyond it,” said Eric Guichard, senior vice president and general manager of AI manufacturing in Silvaco’s TCAD business unit.
Finding an unfamiliar operating region may be exactly what makes AI useful, but it also means engineers have less experimental evidence against which to check its predictions. AI can identify thousands of candidate solutions, leaving engineers to determine which ones justify expensive silicon experiments. Before those recommendations can influence production, manufacturers need to establish where the predictions are reliable and where additional measurements are necessary.
Similar limitations appear in advanced packaging inspection, where a model may distinguish subtle differences in X-ray images without sufficient physical evidence to determine which indicate actual defects. A confident prediction from incomplete information becomes particularly risky when AI is authorized to act on it. An inappropriate adjustment could affect material that continues through several manufacturing steps before the problem is discovered.
When the necessary data doesn’t exist
Semiconductor manufacturers have collected process and equipment data for decades, but much of it remains distributed among separate systems. Equipment logs, inspection records, and engineering databases may contain information about the same product without being connected in ways that make the relationships easy to analyze.
“The growing data challenge was there for us prior to AI,” Fried said. “Five to ten years ago, we really started rolling up our sleeves and building integrated data frameworks for our companies. Those of us who invested earlier and more heavily in building those data frameworks are reaping the benefits of AI now more quickly. Maybe AI has turbocharged this problem, but we’ve had this massive, growing, and complex interwoven web of data for quite some time now.”
Connecting existing data is difficult enough, but manufacturers also face situations where the information needed for a prediction was never measured or isn’t available from the company that generated it. Advanced packaging makes this especially challenging because manufacturing information may originate with several suppliers. A foundry provides the die, another company produces the substrate, and an OSAT performs assembly and test, often without everyone having access to the same manufacturing history.
Some of the most difficult gaps involve physical properties that change during manufacturing. Substrate warpage, thermal stress, and material behavior can affect subsequent operations, particularly as assemblies move through repeated thermal cycles. Those changes aren’t always captured in the available measurements, and a model relying on nominal material properties may miss them entirely.
Manufacturers also need to preserve the assumptions behind the data they exchange. A model developed for one operating range may produce results that appear reasonable when used elsewhere, even though the original calibration no longer applies.
“If the training range, calibration source, or approved use is not carried with the model, it may be good enough for early exploration but not safe for design closu…
AI Brief
Your highlights
Key Takeaways:
AI can identify promising process conditions and recognize manufacturing problems, but those results still require physical validation before they can be trusted in production.
Incomplete manufacturing data and uncertainty about downstream effects make it difficult to give AI systems authority over process adjustments.
Greater manufacturing autonomy will require clear decision boundaries, failure containment, and ways to qualify models that change as they learn.
Semiconductor manufacturing has always depended on engineers making decisions involving complex interactions among process parameters. An etch recipe, for example, may involve combinations of RF power, pressure, gas chemistry, temperature, and timing, each affecting multiple process characteristics. Changing one parameter might improve critical dimension control while creating problems elsewhere. But as devices and manufacturing processes become more complicated, the growing number of variables involved makes those decisions increasingly difficult.
AI-driven manufacturing offers a way to manage some of that complexity, from finding better process conditions to identifying defects and responding to problems on the production line. But how much production responsibility engineers are willing to give those systems remains an open question, and one that drew considerable attention at SEMICON West 2026.
“No matter how good we are as engineers, no matter how disciplined and objective we are, we tend to approach these difficult problems with historical biases,” said David Fried, chief AI officer and corporate vice president of Semiverse Solutions at Lam Research during a conference panel. “A process engineer might look at a set of data and say, ‘I need to move into a lower-pressure, higher-power regime because that’s how we solved this problem 10 years ago.’ But with this massive parameter space, AI can point us in a different direction, into an operating region we might never have considered because we couldn’t see those relationships in the data before.”
Finding a process that keeps working
Finding an unfamiliar operating condition is useful, but determining whether it will hold up in production is another matter. Two recipes may produce similar results during development, yet respond quite differently to changes in incoming materials or equipment conditions. AI can help engineers identify those differences before committing to a particular process.
“It’s not just a question of delivering to a particular performance point. It’s also a question of how stable that particular performance point is, how robust it is,” said Vikas Sharma, senior principal engineer and senior director of pathfinding for Intel Foundry R&D, in a recent presentation. “Minor perturbations in the incoming material or at the tool can lead to a significant drop in performance.”
The scale of the optimization problem helps explain why AI-driven manufacturing is attracting so much interest. Fried described a current Lam Research etcher with a potential recipe space on the order of 10²³ combinations. At Intel, Sharma discussed a process involving 350 parameters and approximately 10⁷⁰⁰ theoretical combinations.
According to Sharma, engineers working with equipment suppliers spent about a year developing a minimally acceptable recipe for that particular tool and process layer, followed by an extended effort to improve it with limited success. After introducing AI-assisted optimization, they began finding better results within a few months, eventually identifying multiple solutions that improved on the original baseline.
The challenge was determining which solutions were worth pursuing. Sharma noted that AI could identify thousands of candidates, but evaluating them still required expensive silicon experiments. Engineers also had to account for nonlinear relationships among parameters, which could change considerably across different operating regions.
“AI is not a magic wand. You can’t use the same AI solution for every problem,” Fried said. “Figuring out whether you’re using full physical simulation, surrogate models, machine learning, deep learning, LLM-based AI, or agentic AI, which of these solutions you’re building in which form to solve which problem and address which workflow, that’s more of what we’re spending our time on than actually building it or implementing it.”
The different approaches serve different purposes, from exploring process conditions to coordinating engineering workflows. Combining them may allow engineers to investigate more possibilities, but it also raises questions about how the results should be validated and whether the assumptions used in one model remain appropriate when its recommendations are passed to another.
When the factory acts on what it sees
Inspection is one of the more established applications of AI in manufacturing, particularly in advanced packaging, where finer pitches and more complicated structures have increased the number of connections that must be examined. Some packages can contain hundreds of thousands of interconnects, each presenting a potential failure point.
Automated optical inspection and X-ray systems collect extensive information about those structures. Machine learning can identify variations that might be difficult to distinguish visually, but manufacturers still need to determine which differences indicate actual defects.
“When we look at X-rays, everybody just sees all shades of gray, but AI can see a thousand shades of gray,” said Ingu Yin Chang, executive vice president at ASE Group. “But the AI only knows if this shade of gray equates to a problem once we say, ‘Hey, this equates to a problem.’ So, you’ve got to have the physical data that says, ‘Okay, that is a problem.’ Then, in the future, they can scan for that shade of gray and that location.”
Once manufacturers determine which image characteristics correspond to actual defects, AI-enabled inspection systems can flag abnormalities during production and trigger predefined equipment responses. At ASE, systems can recognize known abnormalities and trigger equipment responses before additional assemblies are affected.
“We have figured out how to do a lot of automation and use AI to help us identify potential issues, based on the issues that we physically learned through traditional manufacturing methods,” Chang said. “We will know when a wire bond is not bonded correctly, and the machine will immediately stop. There’s constant AI oversight to ask, ‘Are we doing everything correctly? Are we seeing it as we planned to see it?’”
Stopping equipment when a defect is detected can prevent additional losses, but it doesn’t necessarily mean AI is controlling the process independently. The inspection system may identify an abnormal condition while a conventional controller executes a predefined response. Allowing AI to select new bonding parameters and resume production would require considerably more confidence in its decisions.
ASE’s product diversity makes that challenge more difficult. Chang noted that hundreds of package types can pass through the same production lines with different materials, structures, and process sensitivities. A model would have to recognize which product is being manufactured and whether the conditions used during training remain applicable.
An inspection system might correctly identify a defective bond, but changing the process to prevent another failure requires knowing what caused it and whether an adjustment will create problems elsewhere. Manufacturers need considerably more information before allowing AI to make those decisions independently, and much of what’s needed remains difficult to obtain.
Proving the model is right
One of the first obstacles to AI-driven manufacturing is establishing whether an AI recommendation accurately represents the physical process. A model can identify statistical relationships among manufacturing parameters without necessarily capturing the physical mechanisms behind them. That may be sufficient for finding promising operating conditions, but it becomes a greater concern when the system is expected to make decisions that affect production.
Sharma’s Intel example illustrates the problem. Pressure and temperature might dominate the response in one operating region, while voltage or gas flow could become more important elsewhere. A model trained on measurements from one region cannot assume those relationships will hold across another.
“The fundamental principle is that surrogate models are reliable when interpolating within the region represented by their training data and less reliable when extrapolating beyond it,” said Eric Guichard, senior vice president and general manager of AI manufacturing in Silvaco’s TCAD business unit.
Finding an unfamiliar operating region may be exactly what makes AI useful, but it also means engineers have less experimental evidence against which to check its predictions. AI can identify thousands of candidate solutions, leaving engineers to determine which ones justify expensive silicon experiments. Before those recommendations can influence production, manufacturers need to establish where the predictions are reliable and where additional measurements are necessary.
Similar limitations appear in advanced packaging inspection, where a model may distinguish subtle differences in X-ray images without sufficient physical evidence to determine which indicate actual defects. A confident prediction from incomplete information becomes particularly risky when AI is authorized to act on it. An inappropriate adjustment could affect material that continues through several manufacturing steps before the problem is discovered.
When the necessary data doesn’t exist
Semiconductor manufacturers have collected process and equipment data for decades, but much of it remains distributed among separate systems. Equipment logs, inspection records, and engineering databases may contain information about the same product without being connected in ways that make the relationships easy to analyze.
“The growing data challenge was there for us prior to AI,” Fried said. “Five to ten years ago, we really started rolling up our sleeves and building integrated data frameworks for our companies. Those of us who invested earlier and more heavily in building those data frameworks are reaping the benefits of AI now more quickly. Maybe AI has turbocharged this problem, but we’ve had this massive, growing, and complex interwoven web of data for quite some time now.”
Connecting existing data is difficult enough, but manufacturers also face situations where the information needed for a prediction was never measured or isn’t available from the company that generated it. Advanced packaging makes this especially challenging because manufacturing information may originate with several suppliers. A foundry provides the die, another company produces the substrate, and an OSAT performs assembly and test, often without everyone having access to the same manufacturing history.
Some of the most difficult gaps involve physical properties that change during manufacturing. Substrate warpage, thermal stress, and material behavior can affect subsequent operations, particularly as assemblies move through repeated thermal cycles. Those changes aren’t always captured in the available measurements, and a model relying on nominal material properties may miss them entirely.
Manufacturers also need to preserve the assumptions behind the data they exchange. A model developed for one operating range may produce results that appear reasonable when used elsewhere, even though the original calibration no longer applies.
“If the training range, calibration source, or approved use is not carried with the model, it may be good enough for early exploration but not safe for design closure,” said Kenneth Larsen, senior director of product management at Synopsys. “The missing information is not just more data; it is the context that tells the next tool or team how the data should be used.”
As AI systems begin coordinating decisions across multiple process steps, those gaps become more consequential. A recommendation based on one operation may overlook conditions affecting the next, particularly when the model lacks information about the product’s actual physical state. Additional metrology and characterization may help close some of those gaps, but engineers still have to establish whether the available information is sufficient to act on a recommendation.
Who decides what AI can do?
Even with reliable models and better data, manufacturers still have to determine which decisions should be delegated to AI. Existing equipment automation generally operates within established procedures and control limits, but an AI agent capable of evaluating conditions and selecting corrective actions raises questions about who authorizes those decisions and what happens when something goes wrong.
An agent investigating a process problem might need to retrieve inspection results, examine equipment histories, and evaluate possible corrective actions using several engineering tools. Each activity may involve different access permissions, particularly when the information comes from separate facilities or manufacturing partners. Access to those resources also needs to be distinguished from the authority to initiate changes.
“There is a question of data access and data consolidation, and there is also a question of data interpretation,” Sharma said. “But the problem that AI has brought forth as a novel thing, which was not as big an issue before, is governance. When agents are working, how do you ensure that the appropriate privileges for the right data are available to the right agent and accessible to the right person?”
An agent might be authorized to examine process records and recommend an adjustment without having permission to modify a recipe or restart equipment. Those restrictions become harder to enforce when agents coordinate activities across multiple systems, particularly if one tool can initiate actions through another. Access to one system cannot automatically confer authority over every subsequent operation.
Even an authorized adjustment may carry risks. Changing bonding force or temperature, for example, might correct an immediate defect while affecting adjacent connections or downstream reliability. Engineers would have to establish which adjustments can be made within qualified operating ranges, when approval is required, and how the system should respond if a proposed correction falls outside its validated capabilities.
If an AI-directed correction produces an unexpected result, equipment may need to stop while affected material is held for investigation. Manufacturers would also need to establish whether the system should attempt another correction, return control to an engineer, or prevent subsequent operations from proceeding. Those decisions have to be defined before autonomous adjustments are permitted.
Future agentic architectures will likely rely on agents reviewing and verifying one another’s work. That could provide another level of oversight, although the interactions themselves would need to be controlled. Agents relying on the same incomplete information might reinforce one another’s mistakes, and their collective actions could have consequences that none of the individual systems anticipated.
Recent research into agentic AI has demonstrated cases in which systems pursued assigned objectives through actions their operators never intended, including circumventing restrictions or concealing information. [1, 2] Although the research includes controlled experimental scenarios and incidents involving coding agents, it raises questions about how multiple agents might behave when working together.
In manufacturing, a series of individually permitted adjustments could collectively move a process beyond its qualified operating conditions, even if no single action violated its assigned limits. Preventing that would require controls over the combined effects of agent decisions, along with independent verification that the resulting process remains within acceptable limits.
Manufacturing traceability adds another requirement. Engineers investigating a later failure must be able to reconstruct what an agent observed, which alternatives it evaluated, why it selected a particular action, and what material was affected. Those records would need to connect the model’s decisions with equipment settings and manufacturing history.
The amount of authority manufacturers can safely delegate will depend on how reliably AI can operate within those boundaries. A system may be capable of choosing a correction, but it also needs to recognize when the available information is insufficient or when the consequences of an adjustment cannot be predicted with enough confidence.
Maintaining those limits becomes even more complicated when AI models are updated with new production data, potentially changing their behavior after the original operating boundaries have been established.
Qualifying a system that keeps learning
A semiconductor fab may process thousands of wafers between updates to a particular AI model. During that time, equipment conditions change, new material arrives, and engineers collect information that could improve the models’ predictions. Incorporating those results more quickly would be valuable, but manufacturers also need to know whether an updated model will behave as reliably as the one it replaces.
“We are far, far, far away from continuous learning,” Sharma said. “We do episodic learning, once every quarter a model is changed, or once every month a model is changed. But if a factory is running 10,000 wafers, we have 10,000 wafers times whatever number of weeks in between those model changes. How are we using that data in a way that rapidly applies to the learning of that particular model or the agent, or both?”
More frequent updates would introduce additional qualification work. A model might become better at recognizing one process excursion while becoming less reliable under conditions represented in its previous training data. If that model can adjust equipment settings, engineers need to establish what changed and whether the updated behavior remains within qualified limits.
Keeping track of those changes also requires more than identifying which software version was running. Engineers need to preserve the model’s training and validation history, along with the process conditions under which it was qualified. If an AI system changes its behavior during production, investigating a failure weeks later may require reconstructing the model as it existed when a particular lot was processed.
While manufacturers already have procedures for qualifying process changes and software revisions, an AI system that updates itself during production would have to fit within those controls, potentially requiring restrictions on what the model can change and independent checks before revised behavior affects equipment. Changes in manufacturing conditions would also need to be distinguished from changes introduced by retraining the model itself.
The harder question is whether manufacturers can obtain the benefits of continuous learning without having to repeat an extensive qualification process after every update. AI can already help engineers make better manufacturing decisions, but giving it responsibility for those decisions will require evidence that its behavior remains predictable and that its actions can be traced back to validated process conditions.
The path to greater autonomy
Semiconductor manufacturers have spent decades developing processes that can produce consistent results across equipment, materials, and production environments. AI introduces new possibilities for improving those processes, but it also introduces uncertainty about how decisions are made and whether the results can be reproduced. The same standards of process control, qualification, and traceability will have to apply when software begins making decisions that engineers previously handled themselves.
Greater autonomy is likely to emerge gradually, beginning with decisions that can be made within well-characterized operating limits and expanding as manufacturers gain confidence in the systems controlling them. AI may eventually identify a problem, determine an appropriate correction, and verify the result without engineering intervention. Getting there will depend on whether manufacturers can demonstrate that those decisions remain reliable as equipment conditions, materials, and the models themselves change.
References
A. Lynch, J. Hughes, A. Serrano, R. Kirk, and S. R. Bowman, “Agentic Misalignment in Summer 2026,” Anthropic, July 13, 2026. https://alignment.anthropic.com/2026/agentic-misalignment-summer-2026/
OpenAI, “Suppressing Misalignment Monitoring Alerts and Disabling Misalignment Monitoring,” OpenAI Alignment, October 20, 2026. https://alignment.openai.com/misalignment-reports/suppressing-misalignment-monitoring-alerts-and-disabling-misalignment-monitoring/
Related Articles
Shift Left Complicates Fab Data Management
Streaming sensor data creates new challenges for advanced process control.
AI Models Transform Defect Inspection And Review, But Can Fail To Scale
Majority of AI initiatives falter; synthetic data gaining traction due to limited real-world data.
Agentic AI In Chip Manufacturing
What happens when you take the human out of the loop.