Factory Digitalization Ver06. “AI Is Weak with New Products.”- So What Should Manufacturers Do?
- Shigenori Tanaka
- 3月30日
- 読了時間: 3分
Mar 30, 2026
Thank you for reading.
Over the past six years, I have been deeply involved in the digitalization and AI implementation of sand-casting factories. Through this experience, I realized one undeniable truth:
AI has clear limitations.
And when it comes to new products, AI is particularly weak. Today, I would like to explain why and what manufacturers should realistically do about it.
1. AI Has a Fundamental Limitation
- It Cannot Operate Without Past Data
Machine learning models rely on:
Past production records
Past defect histories
Past inspection data
Past process conditions
AI identifies patterns only from historical data.
In other words:
Without past data, AI cannot make any meaningful judgment.
This is not a flaw—it is simply how AI works.
**For why data accuracy matters so much, see my past blogs:
2. But Manufacturing Must Always Produce New Products
This is where the contradiction begins.
Manufacturers constantly face:
New molds
New materials
New processes
New customer specifications
Every month, new products must be launched.
Naturally, new products have:
No production history
No defect history
No inspection data
No optimal-condition records
In short:
New products have zero learning material for AI.
**For the structural limits of factory digitalization, see:
3. There Are Only Two Valid Approaches for New Products
This is the global standard—and exactly what I practiced in real factories.
<Approach 1>
Use Similar Products to Define Internal Parameter Windows
Even for new products, we never start from zero.
We reference:
Material
Wall thickness
Mold design
Cooling conditions
Pouring conditions
Past defect tendencies
From these, we define:
“For this new product, these are the acceptable parameter ranges.”
This is essentially Transfer Learning in the AI world.
<Approach 2>
Ask Process Experts to Create “Generic Pattern‑Based Prescriptions”
Casting, mold, and process experts can derive optimal conditions from theory and experience:
• Cooling drum outlet temperature
• Sand mixing cycle
• AFS
• Active / inactive clay
• Moisture
• Permeability
• Pouring temperature & speed
• Molding compression strength
• Compactability
This becomes the “process prescriptions” for the new product.
4. Conclusion
- AI Works Only on Models Created by Humans
No matter how advanced, AI is nothing more than mathematical optimization.
Without a human‑provided learning model, it cannot function.
Therefore, for new products, AI requires:
1. A similar‑product model (data)
2. A process expert's prescriptions (knowledge)
Without these two pillars, AI simply cannot operate.
5. Over the Past Six Years,
The Question My Customers and I Have Faced Was This:
“How Do We Build the Learning Model That Enables AI?”
My work was never “AI implementation” itself.
It was:
Connecting data sources
Integrating batch‑level production data
Linking inspection results
Converting expert knowledge into structured rules
Turning initial lots into learning datasets
Preventing AI from mislearning
Continuously improving the environment around AI
In essence:
My role was to design the environment in which AI can learn—together with the customer.
Based on the realities of manufacturing, this perspective might offer a useful hint for moving AI operations forward successfully.
Contact
If you are facing challenges in factory digitalization, AI implementation, or PMO leadership in “no one wants to step forward” domains, I would be glad to support quietly and professionally.
Feel free to reach out: info@metricjapan.com
TAGS
コメント