Page 166 - CW E-Magazine (13-5-2025)

Special Report

From real data to synthetic data: Transforming

AI-powered inspection in the chemical industry

Turning the tide on manufacturing Yet traditional inspection systems – DR. WILHELM KLEIN
waste even those using early AI – struggle to E-mail: klein@zetamotion.com
n 2025, sustainability isn’t just a meet the precision required. Accuracy
goal – it’s a metric of resilience. often plateaus at 80%, which falls short ditions, surface textures, and defect
IGlobal industry leaders are under in a fi eld where even minute deviations categories, synthetic datasets make it
mounting pressure to slash emissions, in purity, concentration, or reaction possible to train robust models using
streamline operations, and reduce raw completion can carry signifi cant conse- minimal real-world input.
material waste. Few sectors face this quences.
challenge more urgently than manufac- Instead of waiting weeks to gather
turing, where effi ciency and environ- What’s holding AI back? Dirty data and labelthousands of examples, ins-
mental responsibility are now tightly Behind the scenes, it’s often not pection systems can be developed and
interlinked. the AI models that fail – it’s the data. deployed in a fraction of the time.
According to MIT Technology Review
For chemical manufacturers, the Insights, 57% of manufacturing execu- Training a new quality control spe-
stakes are especially high. Precision is tives identify data quality as their top cialist often involves more than just
non-negotiable. And yet, as Forbes barrier to AI success. In the chemical data – it requires context, patterns, and
reported, up to 40% of production value sector, that number jumps to a stagger- expert guidance. You might show them
is still lost to material waste in this ing 75%. a handful of examples, explain what to
industry. It’s a sobering fi gure – and a look for, and let them build their judg-
call to rethink how quality control is This makes sense: chemical pro- ment through experience. Synthetic
approached. cesses are notoriously complex. Pro- data works in a similar way: it enables
duction logs are incomplete. Variabi- AI systems to learn not from endless
Artifi cial Intelligence (AI) tech- lity is high. And rare defects are, by repetition, but from well-structured,
nologies are stepping up to meet this defi nition, hard to capture. Add to that meaningful scenarios. This approach
challenge. In 2024 alone, over $200-bn legacy systems and fragmented data dramatically accelerates the develop-
was invested in AI for industrial appli- architectures, and you’ve got a recipe ment of accurate, adaptable inspection
cations. From predictive maintenance for bottlenecks. models.
to supply chain forecasting, these tools
are transforming operations. But per- To train a model under these con- From proof-of-concept to production
haps their most meaningful contribu- ditions is like teaching a student with Across the industry, synthetic data
tion to sustainability lies in how they’re outdated, error-fi lled textbooks – and is fast becoming the backbone of scal-
reshaping quality control. expecting straight A’s. Translated to able AI systems. Google and other tech
classic AI, it means “junk in = junk leaders are investing in large-scale
Quality control: The fi rst line of out”. Quality data simply isn’t available synthetic environments for training
defence in many cases. industrial AI. Their message is clear: the
In the chemical industry, quality future of machine learning isn’t just in
control is not just a checkpoint – it’s Synthetic data: Real impact from better algorithms – it’s in better data.
integral to chemical process safety, virtual defects
compliance, and performance. Early Synthetic data is changing the Elon Musk summed it up succinctly:
detection of anomalies prevents batch game. It allows engineers to generate “We’ve now exhausted basically
contamination, reduces the need for re- perfectly labelled, diverse, and con- the cumulative sum of human know-
processing, and ensures product consis- trolled defect scenarios – without ever ledge … in AI training … The only way
tency. These are not just quality issues; halting a production line. to supplement real-world data is with
they are essential for sustainability, synthetic data, where the AI creates
operational effi ciency, and regulatory By simulating inspection data for training data. With synthetic data, AI
integrity. different material types, lighting con- will sort of grade itself and go through

166 Chemical Weekly May 13, 2025

Contents Index to Advertisers Index to Products Advertised