The semiconductor industry faces a problem that goes beyond simulation accuracy itself when attempting to build a digital twin for an advanced package: how can the model remain connected to the package that was actually manufactured, rather than only to the package engineers intended to design? An analysis published by Semiconductor Engineering on September 17, 2026, explains that the answer requires combining design, manufacturing, materials, assembly, testing, and operational data into a single, updateable representation.
The issue is becoming more important as chiplets and heterogeneous integration spread. A chiplet may be characterized, tested, and successful in one package, but reusing it inside a different substrate, with another thermal interface, or within a higher power budget may completely change its electrical, thermal, and mechanical environment. Therefore, the package model cannot stop at the boundaries of the die.
Packaging Is Part of the System Architecture Here
Kenneth Larsen, senior director of product management at Synopsys, distinguishes between a digital twin for a front-end manufacturing process inside a fab with a controlled flow and a digital twin for a package that must represent a system distributed across silicon, the substrate, materials, assembly, manufacturing variation, testing, and reliability data. This does not mean that fab modeling is easy; rather, the nature of the packaging problem is broader in terms of the number of parties and contexts that must be connected.
Teams have effective tools for analyzing electrical, thermal, and mechanical behavior separately, but running these tools side by side does not create an integrated digital twin. The value emerges when the models use the same assumptions and data, then receive manufacturing results so they can change as the package is revised. Otherwise, the twin becomes a late-stage verification tool instead of helping with design decisions before committing to a particular stack-up or a new product.
The Problem Is Not Just a Lack of Data
Joon Ahn, vice president of information technology and factory automation worldwide at Amkor, says that a complete twin requires visibility extending from wafer manufacturing through assembly and testing and then system operation. However, the necessary data is distributed among fabs, substrate and material suppliers, outsourced semiconductor assembly and test companies (OSATs), and customers. An assembly provider may receive information about die dimensions and wafer-test results without detailed maps of power density or localized stress.
Likewise, a substrate may arrive within nominal specifications without actual warpage data for each lot, while a material data sheet does not necessarily reveal how the material behaves inside the final device. Descriptions of the same physical object also differ across disciplines: a chip manufacturer may describe die behavior one way, a substrate supplier another way, while a material supplier provides properties measured under specific conditions. CP Hung, vice president of corporate research and development at ASE, noted that the most difficult handoffs are not always those lacking data, but those that transfer data between different engineering disciplines.
This is further complicated by the fact that the data may include sensitive intellectual property. According to Sanjiv Bhatt, senior director of marketing and business development at Mitsubishi Chemical Group, suppliers and companies retain know-how they do not want to share in full. Therefore, a common file format will not solve the problem of data that was never measured or context that was never transferred.
What Changes When Manufacturing Begins?
The examples in the article show that a nominal model may be accurate but does not necessarily describe the manufactured product. This is particularly clear in photonics, where small changes in dimensions, film thickness, or sidewall shape can affect optical loss or spectral response. Eric Guichard, senior vice president and general manager of Silvaco's TCAD business unit, says that a design optimized according to the drawn geometry does not necessarily represent the manufactured device when large-volume production variations are introduced.
The problem is not limited to photonics. Moisture, the concentration of cleaning materials, incomplete etching, or the properties of a new material may also affect performance. Packaging-material properties are not always fixed numbers either; the coefficient of thermal expansion, modulus of elasticity, and adhesion strength may change with processing, temperature, time, and degree of cure. The model therefore needs to know the conditions under which each value is valid, not merely store the value itself.
Coupled Models Require Continuous Calibration
The effects inside a package do not behave independently. Current generates heat, heat changes resistance, and those changes reshape the power distribution, requiring electrical and thermal calculations to be repeated until convergence. Manufacturing variations may also interact nonlinearly, so that the effect of a group of deviations is greater than, or different from, the sum of their individual effects.
For this reason, the most difficult points of standardization occur at the interfaces between materials, structures, processes, suppliers, and system decisions, rather than necessarily within each individual physical domain. Two models may each be correct within their own domain, but their assumptions may not match when an attempt is made to combine them into a single representation.
Intel's work on hybrid bonding provides an example of the role of measurement in improving a model: experimental deformation data can be entered into a finite-element model, and the refined model can then be used to explore changes that are difficult to test in practice. However, measuring warpage alone is not enough. Ahn suggests that the calibration set should include measurements of joints and bumps, die placement, substrate properties, molding and underfill materials, and thermal-management interface materials, in addition to thermal measurements and electrical tests.
Hung believes that measurement should not occur indiscriminately at every step, but at points where the package state changes, such as bonding, molding, curing, die thinning, and thermal transitions. Calibration also does not necessarily follow a fixed schedule. According to Hanlin Chen, research scientist and team leader at Brewer Science, it should be repeated when a new package structure, substrate, thermal budget, or manufacturing flow moves the material outside the range of conditions on which the model was validated.
Why Does This News Matter?
The editorial conclusion from these facts is that the success of a package digital twin is not measured by its ability to produce an accurate result once, but by its ability to remain valid as the product, materials, and manufacturing change. This places a practical burden on multisupplier supply chains: measurements must be linked to their context, the conditions under which models are valid must be defined, and recalibration must be performed when the package actually changes.
Reduced-order models or artificial-intelligence-based surrogates may accelerate analysis, but they do not eliminate the need to understand their usage limits. As Larsen warns, a fast model, if it loses its context, can become a fast way to make the wrong decision. The open questions remain tied to how much data partners can share, how assumptions can be standardized at interfaces, and how measurement points can be selected to deliver the greatest value without making the manufacturing cycle impractical.