Models need to be synchronized with what manufacturing actually builds.
Key Takeaways:
A chiplet can be fully characterized, tested, and considered known good, but none of that guarantees it will stay that way after it’s placed into a different package. While the chiplet itself may be unchanged, nearly everything around it can move.
A different substrate can alter the mechanical and electrical environment, a different thermal interface can change how heat leaves the device, and a new power envelope can increase how much heat needs to leave in the first place. A thermal path that worked in the first product can cause throttling in the second because the surrounding package and cooling assumptions have changed.
By the time the chiplet is assembled into another package, engineers may be dealing with a different system altogether, which is why package models can’t stop at the die boundary, and why the added complexity makes package digital twins so difficult to achieve. That is the basic problem facing digital twins for advanced packaging. They have to predict how known components will behave after integration and eventually inside a larger system.
“A chiplet may be characterized and validated in one package or product SKU, then reused in a different package stack, substrate, cooling approach, or power envelope,” said Kenneth Larsen, senior director of product management at Synopsys. “The chiplet has not changed, but its electrical, thermal, and mechanical environment has.”
Fab vs. packaging: A different kind of twin
Digital twins are starting to be used in semiconductor manufacturing, but the comparison with front-end fabs needs some clarification. A fab process twin can be enormously complicated, incorporating equipment behavior, process history, and statistical learning. The difference is that much of that complexity remains inside a relatively well-defined manufacturing environment.
“A fab process twin is extremely complex, so I would not frame this as fab being easy and package being hard,” Larsen said. “The difference is the nature of the problem. A fab twin is usually deep within one manufacturing environment, one process technology, and one managed flow of equipment, materials, process data, and statistical learning. A package digital twin has to solve a broader integration problem — how a multi-die system is realized across silicon, substrate, assembly, manufacturing variation, test, and reliability learning.”
That broader integration problem changes what the twin has to represent. In heterogeneous integration, the package is increasingly part of the system architecture, so electrical, thermal and mechanical behavior cannot be optimized independently. Engineers already have effective tools for each of those domains, but they don’t become a package twin simply by running next to one another.
The harder problem is keeping them synchronized around the same physical package as it moves from design to manufacturing. Manufacturing results also need to feed back into those models as the package is revised.
“The package, silicon, substrate, materials, thermal solution, and manufacturing assumptions have to be explored together,” Larsen said. “If they are optimized separately, the digital twin becomes a late-stage validation artifact. That is not where the value is.”
The boundary also becomes less obvious once the package moves onto a board and into its eventual system. Conditions outside the package can change its behavior, which means some of that context has to travel with the model.
“An end-to-end package digital twin requires visibility across the entire lifecycle, from wafer fabrication through assembly, test, and ultimately system operation,” said Joon Ahn, vice president of global factory IT and automation at Amkor. “Some key process information originates upstream in foundry operations or downstream in customer applications.”
Preserving that context would be difficult even if all of the necessary information were sitting in one place. Often, it’s not.
The data hole
From the OSAT perspective, much of the information needed to describe an advanced package comes from elsewhere. Some of it comes from the foundry, some from substrate and material suppliers, and some is generated only after assembly or test. Even equipment inside the same manufacturing flow may provide different types of data depending on its age and supplier.
“For OSATs, an end-to-end digital twin is not a single model, but a connected ecosystem of process, equipment, production, supply chain, and traceability data,” said Ahn. “Packaging performance is influenced by interactions among multiple process steps. Some key process information originates upstream in foundry operations or downstream in customer applications.”
What an OSAT receives can be quite different from what a predictive model ideally contains. Incoming die data may include critical dimensions and wafer-sort results, but not detailed power-density or local stress maps. A substrate may arrive within nominal specifications without showing actual warpage across the production lot, while material data sheets may not capture material-specific behavior in the end device.
“The downstream analysis can then look precise while missing the manufacturing reality,” Larsen said. “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.”
That context can be lost even when plenty of data survives the handoff. The problem is that different parts of the supply chain may describe the same physical system in different ways.
“The hardest handoffs aren’t necessarily where we have the least data. They’re usually where the data has to move between very different engineering domains,” said CP Hung, vice president of corporate R&D at ASE. “A foundry may describe die behavior one way, a substrate supplier another, and a materials company may provide properties measured under a very specific set of conditions. Then the OSAT has to bring all of that together and understand how it will behave during assembly.”
That context becomes even harder to preserve when the information crosses company boundaries. The material supplier knows things about the chemistry that the equipment company doesn’t. The equipment company knows details of the process integration that may not be available to the OSAT or fab. Much of that knowledge also represents intellectual property.
“We all have our own data and everybody wants to protect intellectual property,” said Sanjiv Bhatt, senior director of marketing and business development at Mitsubishi Chemical Group. “I’ve been in this industry all my life, and when we started we had a different way of collaborating. We were a lot more open as an industry, as an ecosystem.”
For a package digital twin, this makes the data problem partly technical and partly organizational. A common file format can help move information between tools, but it can’t provide process data that was never shared, a material property that was never characterized, or an excursion that occurred farther upstream and was not attached to the incoming die. The model can be incomplete before the first coupled simulation is run.
Manufacturing gets a vote
Even complete nominal data only describes the package engineers intended to fabricate. Manufacturing still gets a vote, and it doesn’t always vote for the nominal geometry. The difference is particularly visible in photonics, where relatively small shifts in geometry can materially alter device behavior. Critical dimensions, film thickness, or sidewall profile can move enough during processing to change optical loss or wavelength response without anything being obviously “broken.”
“The biggest disconnect is that photonic devices are typically designed and optimized using idealized geometries, while manufactured devices are governed by process-induced variability,” said Eric Guichard, senior vice president and general manager of Silvaco’s TCAD business unit. “A photonic integrated circuit may demonstrate excellent performance in simulation or in a carefully controlled research run, but high-volume manufacturing introduces variations in dimensions, profiles, and materials that directly affect optical behavior. As a result, a design optimized for the ‘as-drawn’ geometry often does not represent the ‘as-manufactured’ device.'”
Fig. 1: GPU-powered 3D FDTD (finite-difference time domain) simulation shows variations in the resulting electric field due to manufacturing variations. Source: Silvaco
The same problem extends beyond photonics. The model may describe the nominal structure correctly while missing a manufacturing effect that appeared secondary until the hardware is built.
“Subtle nuances don’t show up in the model,” said Brad Booth, CEO of NLM Photonics. “We’ve kind of done this with ASIC design tools. We can simulate stuff first, and what you get out of the back end is usually very close to what you represented in a software model. When you start bringing in things that haven’t had that level of lifetime or qualification, or if you’re pushing the bleeding edge, it’s a little harder because now you’re introducing subtleties, and those subtleties can have massive impacts.”
Fine process details such as humidity levels, changes in cleaning chemistry concentration, or a slight under-etch become model inputs if they change performance. New materials add another complication because engineers need to know not only their nominal properties, but how those properties change with dimension, environment (especially temperature), and time.
“A datasheet gives you a snapshot — CTE, modulus, adhesion, viscosity, measured under one set of conditions at one point in time,” said Hanlin Chen, research scientist and team lead at Brewer Science. “Mechanical and thermal properties shift with the degree of curing. Crosslink density may continue to evolve after processing, and adhesion strength depends heavily 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.”
That creates a moving target even when the material name and formulation have not changed. A value that was accurate when the model was built may become less representative after cure, thermal exposure, or assembly. So the problem is not simply about collecting more material properties. Engineers also must know when those properties apply.
“When you introduce any new material into that, it just explodes the number of knobs and dials, because you’ve got to know how that material performs in all these different environments and what its variance is over its lifetime,” Booth said. “You can model to a certain level and say, ‘This is the ballpark I should be in.’ But then you actually have to build the parts because you have to have some way to correlate back to that model.”
Building the parts settles one question and immediately creates another. Once new materials are assembled, their variations stop behaving independently. At that point, building the twin is no longer mainly a matter of adding better inputs. It also involves changes caused by the thermal and mechanical environment.
When one problem becomes another
That interaction is where package digital twins begin to depart most sharply from a collection of conventional simulations. Engineers may model thermal, mechanical, and electrical behavior separately for expediency, but the physical package doesn’t respect those boundaries.
Once those effects begin feeding back into one another, separate models become harder to keep separate. Power is a straightforward example. Current creates heat, heat changes resistance, and those electrical changes alter the power being dissipated. The new power distribution changes the thermal picture again, so engineers may have to iterate between electrical and thermal models until the results converge.
The same problem can happen in process variation, where one change can alter the effect of another rather than simply adding to it.
“A key observation is that these variables rarely act independently,” said Silvaco’s Guichard. “Yield is often limited by the combined and nonlinear interaction of multiple process variations.”
This is also where standardization gets difficult. Models for heat transfer or electrical behavior may be well understood on their own, but the assumptions used at the interfaces are less portable.
“The least-standardized models are often at the interfaces, not in the isolated physics domains,” Synopsys’ Larsen said. “The difficult areas are where materials, structures, processes, suppliers, and system-level decisions meet.”
A substrate supplier and an assembly house may both have technically valid models, for example, while describing different parts of the same physical behavior under different assumptions. The problem is not necessarily that either model is wrong but they may not mesh cleanly enough to produce a common representation of the package.
Calibration never ends
Eventually, the physical package gets a chance to argue with the model. When it does, the measurements become part of the twin itself rather than a final check. Intel’s hybrid-bonding work offers a useful example. Experimental distortion data is fed back into the finite-element model, which engineers can then refine before using it to explore changes that would be impractical to test experimentally.
The goal is to narrow the range of conditions engineers need to explore to reduce distortion and eventually understand how sensitive that behavior is to changes in process and materials, because one measurement rarely tells the whole story. Warpage is an obvious calibration target for advanced packaging, but meeting a warpage target doesn’t show that everything else in the package is behaving as expected.
The harder question is deciding when another measurement is actually needed. Following every variable through every process step would quickly become impractical, but some manufacturing steps alter the package’s physical state enough that the model needs another reality check.
“I don’t think the answer is to measure everything at every step,” said ASE’s Hung. “It’s really about measuring at the points where the package changes state.” Those transitions can include bonding, molding, curing, thinning, and thermal transitions.
For this reason, modeling has to consider all package interfaces and test results. “For a package model, warpage is the obvious measurement, but it is not enough,” said Amkor’s Ahn. “The calibration set has to include bump and interconnect metrology, die placement, substrate characterization, underfill/mold/TIM measurements, thermal characterization measurements, and electrical testing.”
Model development and physical development therefore cannot be separated when the technology itself is new.
“You’ve got to build that twin at the same time you’re building the devices, so you have some correlation,” NLM’s Booth said. “We’re having to literally do both of those in lockstep, because as much as I can go and create this thing and say, ‘Here’s this great circuit,’ until I’ve actually built one and shown that I can build it and it equates to what was in the digital twin, I can’t really say that my digital twin is an accurate representation of the end product.”
Nor does a successful calibration remain valid indefinitely. A change in the package, materials, or process can push the next application beyond the conditions used to validate the original model.
“There’s no fixed interval; it’s driven by how far the new application sits from the conditions of the original model,” said Brewer Science’s Chen. “A new package architecture, a different substrate, a shift in thermal budget, or a change in process flow can all move a material outside its validated envelope. Our approach is to recalibrate whenever the customer’s application introduces a meaningful change in those variables, rather than on a calendar schedule.”
Calibration also has its limits. It is only useful as long as the conditions that made the model valid remain attached to it. That becomes particularly important with reduced-order and AI surrogate models, where speed depends on using a simplified representation without losing track of the conditions under which it remains trustworthy.
“Without that context, a fast model can become a fast way to make the wrong decision,” Larsen said.
Conclusion: Keeping the twin attached to reality
The difficulty in building a package digital twin goes beyond creating accurate electrical, thermal, or mechanical models. The challenge is keeping those models attached to the same physical package as manufacturing changes it and new information becomes available. Measurements from fabrication, assembly, and test have to feed back into the representation, along with enough context to show what those measurements actually describe.
That is particularly difficult in advanced packaging because there may never be a point when everything is completely known. Components come from different suppliers, materials change during processing, and assembly introduces new stresses and interactions. The twin has to evolve along with the package or risk becoming a precise model of something that no longer exists.
“A useful package twin is not just a simulation sitting next to the hardware,” Larsen said. “The built package should improve the model, and the model should let teams ask what will change before they commit to the next design, package stack, or product SKU.”
That may be the real measure of success. The twin does not have to predict the physical package perfectly once. It has to remain useful as the physical package changes.
Related Articles:
Digital Twins For Packaging: Bridging Design, Fab, Test, And Reliability
Packaging moves toward predictive control, but trust and validation still lag.
Digital Twins Find Their Footing In IC Manufacturing
Technology will speed time to yield and add efficiency, but standards are needed for it to live up to its potential.
How Long Does A Measurement Remain Valid?
Measurements can be misleading after the chip, the test path, or the assumptions behind the result change.
Can Fine-Pitch Hybrid Bonding Go High Volume?
Hybrid bonding already works in production, but finer-pitch die-to-wafer integration must preserve fab-level surface and alignment control at back-end volumes and costs.
Leave a Reply