半导体模型能隐藏多少信息？

Key Takeaways:

• Advanced packaging pushes process behavior across company lines, so what one supplier hides now shapes another company’s prediction.
• Encryption and abstraction protect the recipe cleanly, but neither tells the receiving engineer where the model stops being valid.
• A model can be accurate and still be wrong for the question asked, making its envelope as much a deliverable as the model.

The goal seems simple — share the model without sharing the process. A materials supplier shouldn’t have to reveal a proprietary formulation just so a customer can predict how it will behave in a package, and a foundry shouldn’t have to hand over its recipe to provide a useful model of a chiplet. But advanced packaging is making that separation much harder to maintain.

For years, engineers moved much of what they needed between companies through datasheets, design rules, and qualification specs, while the manufacturing knowledge stayed with the supplier. That arrangement held as long as the pieces behaved independently, but what one company keeps inside its process now shapes how another company’s part of the package performs once everything is assembled. The process may still belong to the supplier, but its consequences no longer end at the factory gate.

“There’s been some initial work on this challenge, but how do you insert a membrane between the companies so they can share a certain amount of data, but not all,” said Greg Prewitt, director of product management for Exensio Test Operations at PDF Solutions. “Do you obfuscate it? Do you normalize it? How much do you share so both teams can solve a problem without seeing any of the IP that’s proprietary to one company or another?”

The difficulty is deciding what that membrane should let through. Hide too little and the supplier gives away process knowledge it spent years developing. Hide too much and the model may lose the behavior, variation, or limits that the engineer on the other side needs to make a useful prediction. That leaves a harder question — how much can disappear inside it before it starts hiding something the prediction actually needs?

Materials show the problem in its simplest form. Engineers may need to know how a material changes with temperature, whether a substrate will warp enough to affect assembly, or how process variation propagates into electrical performance without needing the proprietary details that produced those behaviors. A customer doesn’t need a material company’s formulation to model how one of its materials will behave, but a few static numbers on a datasheet may not be enough either.

“Although proprietary formulations cannot be shared, material characterization across the range of conditions relevant to the intended application can provide a more comprehensive representation of material behavior, allowing models to better reflect real-world conditions and improve prediction accuracy,” said Hanlin Chen, senior research associate and team lead at Brewer Science. “A digital twin that uses single fixed values for material inputs will always underrepresent real-world risk.”

The data handoff challenge
The need for more detailed models is growing because an advanced package brings together pieces whose engineering histories accumulated in different places. They may pass through several companies before the finished package reaches the system, but by then no single company necessarily has the whole history.

“Advanced packaging is enabled by a highly collaborative supply chain,” said Joon Ahn, vice president of global factory IT and automation at Amkor. “Dies, substrates, materials, and design information are contributed by multiple partners. Each participant owns unique expertise and data associated with their portion of the value chain.”

That distributed ownership becomes a problem once the design flow needs manufacturing information before everything has been built. The process can remain inside one company, but enough of its behavior has to become usable somewhere else.

What reaches the next engineer is usually enough to describe the part as designed, but not necessarily enough to explain how it will behave after assembly. Nominal dimensions and basic electrical or material properties are relatively easy to pass along. Production variation and the history behind those numbers are much less likely to make the trip.

“I don’t think we need complete data transparency to create a connected digital thread,” said CP Hung, vice president of corporate R&D at ASE. “What we really need is a trusted way to share enough information for the next engineering decision to be made.”

That doesn’t mean every upstream detail belongs in the model. Some can disappear without changing the prediction, while others matter only when something elsewhere in the package exposes their effect. For example, a chiplet’s thermal response may be enough for one analysis, but process variation can change that response in a way the downstream model can’t see.

The question therefore isn’t simply how much information can be shared. It’s which information still matters after the process itself has been hidden. A model that preserves the right behavior can protect the recipe underneath it. Hide the wrong thing, however, and the uncertainty disappears from the model even though it never disappeared from the package.

How much physics can a model hide?
Reduced-order and surrogate models solve a similar computational problem in different ways. A reduced-order model simplifies a higher-fidelity physical representation, while a surrogate approximates behavior across a sampled design or process space. Both make analysis practical when running the underlying model every time would take too long, but they don’t become more trustworthy because they run faster.

“The surrogate model can be trusted if the simulated data used to create the surrogate model is accurate,” said Eric Guichard, senior vice president and general manager of AI Manufacturing at Silvaco. “For the simulation data to be accurate and predictive, the physical models used to generate them must be calibrated. The calibration can only be done against experimental data. Those experimental data are usually available and not proprietary. Physical model calibration is only one part of the equation. The second part is the knowledge of the detailed process recipes.”

The problem splits into two very different kinds of information. Experimental measurements can provide the evidence needed to calibrate a model without revealing much about the process itself, but building a predictive representation may still require details the company would never hand directly to a customer.

What makes that workable is an asymmetry between building the model and using it. The organization creating the model may need to know enough about the process to establish why it behaves the way it does. The engineer consuming the model may only need access to that behavior. A substrate model, for example, might expose thickness, material response, and the effects of process variation without revealing how the supplier achieves those characteristics in manufacturing.

“If only some key knobs are exposed to the user of the surrogate model, we can reasonably believe that the proprietary process details are preserved,” Guichard said.

The recipe doesn’t disappear. It moves behind the model boundary. But that only relocates the disclosure problem rather than solving it, because the model builder still needs what the supplier is trying to protect.

“Availability of detailed process recipes is actually the challenge,” Guichard added.

Encryption protects proprietary data
Encryption creates a way to separate useful behavior from proprietary implementation. Synopsys’ HFSS (High Frequency Structure Simulator) supports encrypted electromagnetic models that a component supplier can share for use in a customer’s simulation without exposing the underlying geometry and material information.

“These encrypted models solve a significant but non-technical hurdle — how to share simulation models of an electronic part with your customers without revealing your trade secrets,” said Lang Lin, product management director at Synopsys.

Encryption handles the disclosure problem neatly. The engineer receiving the model can use it without gaining access to what’s sealed inside. What encryption can’t tell that engineer is whether the model is still valid for the problem being asked.

“That has always been one of the key concerns about simulation, in general,” Lin said. “How well does it actually match reality?”

A model can be secure and still be wrong for the conditions in which it’s being used. If it was built around one package or process window, the result may remain numerically precise long after the design has moved beyond those conditions. Encryption can make that harder to catch because the user can’t look inside and see where the original assumptions stop holding. The IP can stay perfectly protected while the engineering decision doesn’t.

Model sharing therefore needs something outside the encrypted model itself. The recipient may not need to know how the model works internally, but they still need to know the conditions under which it can be trusted.

When the model stops working
Materials make that problem easier to see. A datasheet can be perfectly accurate and still describe only one moment in a material’s life. Once that material is cured, assembled, and put into service, the properties used to build the model may start to move.

“A datasheet gives you a snapshot, CTE, modulus, adhesion, viscosity, measured under one set of conditions at one point in time,” said Brewer’s Chen. “Mechanical and thermal properties shift with degree of curing. Crosslink density may continue to evolve after processing, and adhesion strength is highly dependent on the specific bonding interface. CTE also varies, particularly in the temperature region near the polymer’s Tg (glass transition temperature). Treating any of these as a single static number causes a lot of simulation-to-silicon gaps.”

The supplier doesn’t have to reveal the formulation to account for that behavior. What has to survive the handoff is enough information to show how the material changes under the conditions the package is likely to see. A nominal value may be enough at one point in the design space but misleading somewhere else.

Surrogate models make the same validity problem particularly visible. They can be highly accurate inside the region represented by their training data and still return a plausible-looking answer after the user has wandered outside it. The model hasn’t necessarily failed at that point. It’s being asked a question its underlying data was never meant to answer.

“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,” Guichard said.

The difficulty is that nothing in the output necessarily flags the crossing. A number returned from outside the training region can look exactly like a number returned from inside it.

A useful black box therefore needs a visible envelope around it. The recipe can remain hidden, but the conditions that establish where the model applies can’t. The engineer using it needs to know when the question has moved outside the territory the model actually knows. The model builder has to make that territory visible, and the user still has to respect it.

The interfaces add complexity
The problem gets harder at the interfaces because the IP line often falls in the same place where two independently built models have to begin describing the same physical behavior.

A material, substrate, or die model may be perfectly valid for the conditions its supplier intended. The trouble starts when those models are joined and the assumptions underneath them don’t line up. One may treat a temperature, geometry, or other condition as fixed while another allows it to move.

“The challenge is that the data often comes in different formats, at different levels of detail, and with different assumptions behind it,” said ASE’s Hung. “A material property, for example, only really means something if we know how it changes with temperature, moisture, time, and process history. A component model may also work perfectly well on its own but not tell us enough about what happens once it’s integrated into the package.”

Those assumptions become harder to separate once the package is assembled. Heat generated in the die moves through the structures around it, while the resulting temperature changes material behavior and mechanical stress elsewhere in the assembly. The models may have been created separately, but the package doesn’t operate that way.

“The real challenge isn’t simply moving data from one company to another,” Hung said. “It’s making sure the physical meaning of that data doesn’t get lost along the way.”

Thermal analysis shows how quickly the information problem follows. Substrate and interposer data may exist upstream, but the downstream design flow needs it in a form that can be combined with the rest of the package.

“When you’re talking about thermal, it’s not just the dies,” said Amlendu Shekhar Choubey, senior director of product management at Synopsys. “This is an integrated problem. You need the whole stack.”

But having all the pieces still doesn’t guarantee they fit. One supplier may characterize behavior under one set of conditions while the model on the other side of the interface assumes another. Even the same nominal parameter can mean something different if it was measured at another temperature, derived from a different process state, or intended for a different range of operation.

The harder problem isn’t getting more data across the interface. The handoff has to preserve enough of the assumptions behind the data for the next model to know how to use it, or every model can be right on its own while the package they describe together is wrong.

What models need to carry
The model itself is becoming a kind of engineering contract between companies. One side doesn’t promise to expose how a process works. It promises to describe enough of what that process does for another engineer to use the result within an agreed set of conditions. The formulation, recipe, or implementation can stay hidden as long as the model exposes the behavior the next engineer actually needs. The receiving engineer still needs to know what kind of promise is being made.

“Uncertainty and model validity need to move with the model,” said Kenneth Larsen, senior director of 3D-IC product management at Synopsys. “It is not enough to exchange geometry, material properties, or simulation results. Each model also needs to carry the context that determines whether it can be trusted — where the data came from, how it was measured or calibrated, what assumptions were used, what range it is valid over, and what decisions it is intended to support.”

That context matters most when the model is abstracted or encrypted. A model useful for architecture exploration may not be trustworthy for final closure, and a surrogate built inside one process window doesn’t automatically stay predictive after the package changes. The output alone gives the engineer no indication that the original promise no longer applies.

The practical answer isn’t to demand that everything become visible. It’s to make the model’s useful behavior and its limits visible enough that the company receiving it can decide whether it applies.

Supplier boundaries won’t disappear, and neither will the competitive value of the processes behind them. In many cases those proprietary details are exactly why a customer selected the supplier in the first place. What’s changing is how much engineering information can stop at the boundary with them. A traditional black box could hide both the process and most of its behavior because the customer mainly needed a specification. A model shared into a tightly coupled package has to conceal the recipe while leaving enough of its consequences visible for someone else to make a defensible decision.

Conclusion
For years, the black box was a workable bargain. A supplier could protect how something was made, the customer could design around a specification, and enough margin usually existed between the two to absorb what neither side could see. Advanced packaging is narrowing that margin because the behavior of one piece increasingly depends on what happened to another before the two ever met.

That doesn’t make proprietary process knowledge less valuable. If anything, the process may become more valuable as packages demand materials, assembly methods, and manufacturing windows tuned to a particular architecture. But the model now has to carry something the old specification often didn’t — enough evidence about what that hidden process does, where that behavior changes, and when the representation should no longer be trusted.

The companies that solve this best probably won’t be the ones that share the most. They’ll be the ones that can turn proprietary process knowledge into models other engineers can use without being able to reconstruct what’s underneath. The recipe can stay private, but the consequences cannot.

Related Articles
AI Is Rewriting The IP Playbook
As the semiconductor ecosystem pivots to AI, it is transforming how IP is created, verified, managed, and sold.

How Long Does A Measurement Remain Valid?
Measurements can be misleading after the chip, the test path, or the assumptions behind the result change.

An AI Model Fit For Purpose
Using models outside the context in which they were created for can result in misleading information.