When You Knows but Can’t Act: The Hidden Hurdle in Semiconductor Smart Manufacturing

Taiwan's Semiconductor Edge Creates Unprecedented Pressure

Taiwan’s leadership in semiconductors is the result of decades of investment and deep technical expertise. However, as the industry advances toward smaller nodes, complexity increases significantly. Parameter interactions become more sensitive, product launch cycles accelerate, and process windows narrow. The traditional reliance on experience-based optimization can no longer keep up with these demands.

When you factor in high measurement costs and delayed feedback from metrology tools, the pressure on frontline teams becomes immense. The urgency to improve is constant, yet improvement cycles remain slow and cumbersome. While some companies respond by expanding centralized data teams or investing in massive data lakes, these top-down solutions often overlook the most critical resource: the process knowledge of engineers.

See how NEXCOM and Profet AI's consulting services delivered a solution to empower process engineers, further expanding their expertise with AutoML software and AI computing systems.

Challenge

In semiconductor manufacturing, the gap between identifying a problem and resolving it is often measured in weeks, not hours. The crucial process—virtual thickness measurement for CVD (chemical vapor deposition)—may not be the most glamorous application of AI, but it highlights a fundamental truth about the current state of smart manufacturing.

For years, the conversation around AI in factories has centered on technological feasibility. However, in practice, the real bottleneck isn’t the technology; it’s access. The process engineers who understand the problems best—like the subtle interactions between pressure, flow, and temperature—are often the least equipped to build the data models needed to solve them.

We witnessed this firsthand at a major semiconductor company. Despite having a dedicated team of data scientists, most process improvement ideas from the engineering team never made it to the top of an already long priority list. Engineers were eager to solve these problems themselves, but the tools were too fragmented, and the modeling process felt inaccessible. This isn't a data problem; it's a usability problem.

Virtual Metrology for CVD

In this context, the company introduced Profet AI’s AutoML platform with a simple goal: to let process engineers build and deploy their own predictive models without waiting in line for data scientists.

CVD is a precision-heavy and time-consuming step where film thickness and uniformity are paramount. The traditional workflow involves running trial batches, relying on senior engineers to tune parameters, and waiting for offline inspections to validate results. This feedback loop is slow, and by the time a deviation is found, corrective actions often come too late.

Using the platform, engineers uploaded their historical process data including pressure, gas flow, temperature, timing, and more as the dataset. Within a few hours, they had built and validated models capable of accurately predicting film thickness and uniformity.

Integrated AI Computing System

Profet AI’s AutoML platform is installed in the NEXCOM's TT 300-A3Q, a powerful industrial PC with a PCIe x16 expansion slot for advanced GPU support. Designed to support AI model training and inference, the system is powered by 12th/13th Gen Intel^® Core™ processors and accelerates workloads with Intel^® OpenVINO and Intel^® Deep Learning Boost, making it a perfect companion for the AutoML software.

Comparing to other high-performance products, the NEXCOM TT 300-A3Q's compact size could seamlessly integrate into the limited space, demonstrating its remarkable performance even in the challenging conditions of high temperatures and humidity. The device operates efficiently within a wide temperature spectrum of -5°C to 55°C and a humidity range of 10% to 95%. With all I/Os at front, it is designed for easy maintenance and installation.

The system also supports AI deployments across the factory ecosystem, from edge DAQ (Data Acquisition) to MoM (Manufacturing Operations Management) system, and ultimately integrates with enterprise systems like ERP and MES. AI analyzes incoming operational data from field equipment/PLCs and powers insights in a centralized “Enterprise War Room,” enabling real-time monitoring and KPI-driven decision-making across many AI-enabled modules such as production line monitoring, energy management, and cybersecurity.

In short, NEXCOM TT 300-A3Q transforms traditional factory operations into AI-powered smart manufacturing hubs, enhancing automation, responsiveness, and operational intelligence.

The Numbers: Less Waiting, More Engineering

The impact wasn’t in adding more AI—it was in eliminating the wait. The results shifted the team’s entire workflow:

• Pre-production parameter setup, which used to take 4 hours, was cut in half to 2.

• Quality predictions became available in near real-time, eliminating the delay from physical inspections.

• Physical measurement costs plummeted by an estimated 70%.

• Most importantly, the overall process yield improved by approximately 2%.

The change wasn’t just about gaining access to a new tool—it was about empowering engineers to act directly on their own process expertise.

From a One-Off Project to a Repeatable System

This approach also addresses a common failure point for AI projects: successful models that remain "black boxes" and cannot be scaled because they weren't properly documented.

With the AutoML platform, every step from data selection to model tuning and performance validation, is automatically logged in the platform. This creates a structured, transparent record that allowed future teams to revisit, reuse, and build upon prior work. What starts as a one-off success could now become a repeatable and scalable process.

The True Shift: Empowering Engineers, Not Replacing Them

The core challenge in smart manufacturing has never been a lack of data or a shortage of problems to solve. It has always been a bottleneck of usability.

When process engineers are empowered to test their own ideas and build models directly, companies can finally unlock the hidden value in the vast data they already collect. Profet AI's platform and NEXCOM's IPC didn't just provide a tool; it introduced a brand new, more intuitive way of working. When those closest to the problem are empowered to find the solution, AI moves beyond a buzzword on a presentation slide and becomes a practical part of the daily toolkit.

The future of industrial AI isn’t about making engineers into data scientists—it’s about making data science accessible to engineers.

About NEXCOM: https://www.nexcom.com/index.html