Modified Engineering Plastics

The AI knowledge base for modified engineering plastics plants

A compounding plant runs on changeovers. This week it is PA6 with 30% glass fibre, next week a flame-retardant PC/ABS, then a high-impact PBT for a customer who needs it yesterday. The formula lives in one engineer's spreadsheet, the process window on a paper card at the line, the UL yellow card in somebody's inbox. When that engineer leaves, half the knowledge leaves with them.

RatioMax brings those scattered files into one privately deployed knowledge base. Formula BOMs, process cards, test reports, yellow cards and customer-specific requirements are kept in separate spaces, retrieved with hybrid vector and keyword search, then re-ranked. Every answer links back to the exact passage it came from. Nothing leaves the plant network, the formulas stay where they belong, and a new hire can look things up without interrupting a senior engineer.

Where it breaks

  1. Formulas and process cards scattered

    One grade's formula sits in R&D's spreadsheet, its process window on a paper card at the line, the customer's special requirements in a planner's inbox. The three versions drift apart, and the mismatch only shows up during a trial run, when the answer is to ask the senior operator what he did last time.

  2. Yellow cards, TDS and grades disagree

    A customer asks whether a grade passes UL 94 V-0. The engineer has to open the yellow card, the TDS, and a test report from three years ago, then confirm it is the same flame-retardant system. Every plant names grades its own way, and one formula revision breaks the trail.

  3. New hires learn by shadowing one expert

    Screw configurations, barrel temperature profiles, vacuum level, strand cooling and pelletiser speed are mostly found by trial. That knowledge sits in one senior engineer's head. Ask why a compound needs side feeding and the answer is often 'we have always done it that way'. When he retires, the settings remain but the reasoning does not.

  4. Slow traceability, and no public cloud

    A customer complains about black specks in a lot. Quality traces the batch through raw material records, the process settings used that day, and whether other grades in the same window showed the defect. Records are split across MES, paper route cards and inspection sheets. Bringing in an AI tool feels like handing over the formulas.

How RatioMax responds

  • Private knowledge base, isolated by space

    Formula BOMs, process cards, test reports and yellow cards live in separate spaces with separate permissions, so R&D material stays out of the shop floor view. The whole stack runs inside your plant on Docker Compose or Kubernetes, and each knowledge base can bind its own private OSS or S3 bucket. No data leaves the intranet.

    Platform capability: On-premise, space isolation

  • Compliance questions answered with sources

    Yellow cards, TDS documents, internal test reports and customer technical agreements go into the same index. Hybrid retrieval merges vector and keyword hits, then re-ranks them, so a question about UL 94 V-0 comes back with the exact line on the yellow card and the glow-wire or CTI report paragraph behind it.

    Platform capability: RAG with citations

  • Turning trial-and-error know-how into wiki

    Feed existing process cards, trial records and defect reports to the agent and it drafts cross-linked Markdown pages plus a knowledge graph: which screw configuration suits which grade, why side feeding is used, what black-speck issues showed up before. Pages stay editable, with snapshots, line-level diff and rollback when someone gets it wrong.

    Platform capability: Auto-generated wiki

  • Batch traceability and plant data queries

    A dedicated data-analysis agent writes SQL from plain questions against production, quality and inventory data: what were the zone temperatures on this PA66-GF30 batch, how many grade changeovers stopped the line last month. No waiting on an IT ticket. Each answer keeps its query trace and can be replayed step by step.

    Platform capability: NL2SQL queries

Deployment scenarios

  • A customer asks about a grade on the spot

    A sales engineer at the customer's site is asked for the glow-wire temperature of a flame-retardant PC/ABS. She opens the website widget, asks, and gets the yellow card line plus the internal test report, with the recommended processing window. Answered on the spot.

  • Cleaning checks at a grade changeover

    Switching from a dark flame-retardant compound to a natural reinforced one, the operator follows the cleaning SOP in the knowledge base: pull the screw, purge the barrel, check first parts for black specks and colour difference ΔE.

  • Prepping a new grade before the first trial

    Before building a 30% glass-filled, UL 94 V-0 PA66 grade, the formulator asks the base for how similar grades were built, which flame-retardant package was used, and what blooming or white-core issues showed up. Then he picks screw configurations and temperature windows to trial.

  • Chasing black specks after a complaint

    Several grades report black specks in the same week. Quality asks the data agent one question: which masterbatch lot, which machine, which screw set were these batches run on. Then the maintenance and cleaning history for those machines narrows it to a single changeover.

Terms and knowledge entities

Melt flow index (MFI)
The mass of polymer that extrudes through a standard capillary die in ten minutes at a set temperature and load. Higher MFI means lower molecular weight, easier flow and generally lower impact strength.
Twin-screw extrusion compounding
A compounding process where two co-rotating screws melt, mix, devolatilise, strand and pelletise resin with glass fibre, fillers and additives. The screw configuration sets the shear history and residence time the compound sees.
Glass-fibre reinforcement
Short glass fibres raise tensile and flexural modulus and HDT while cutting mould shrinkage. Fibre loading and fibre-matrix coupling decide how much reinforcement you get and how visible the fibres are.
UL 94 flame rating
A UL 94 vertical-burning rating for plastics. V-0 is the stricter of the common vertical grades: short afterflame, and no burning drips that ignite the cotton below.
Heat deflection temperature (HDT)
The temperature at which a loaded standard specimen deflects by a set amount, used to judge dimensional stability at temperature and usually quoted alongside Vicat softening point.
Moisture content and drying
Polar polymers such as PA, PC and PBT absorb moisture, so they must be dried before moulding. Too much water causes hydrolytic degradation, showing up as silver streaks, bubbles and lower impact strength.
Grade management
Tying each grade to one identifier: base resin, filler or reinforcement, flame-retardant system, formula revision, yellow card, TDS and customer approval status all point to the same record.
Grade changeover and purging
The routine between two grades: pull and clean the screw, purge the barrel, then check the first parts, so dark, flame-retardant or highly filled residue does not carry over into the next compound.
UL yellow card
A public data card UL publishes for a qualified material, listing the grade designation, recognised colours, flame rating and electrical and mechanical properties that customers and certifiers rely on.

These definitions are also exposed structurally so search engines and AI engines can read and cite the domain concepts accurately.

Related deployment

Guangdong Zhongsu New Materials Co., Ltd.

Stock code 301686 · National “Little Giant” specialised enterprise

Founded in 2009, focused on R&D, production and sales of modified engineering plastics, with three production bases in Dongguan Chang’an, Ji’an (Jiangxi) and Dongguan Shuixiang.

Solution: In the Hongmei plant’s AI-powered APS scheduling project, RatioMax delivers the enterprise knowledge base and APS data Q&A, fusing RAG with NL2SQL.

  • ≤30 min Planning time (from 4–8 h per run)
  • 80%+ Faster planning-data queries
  • 30–45% Fewer grade changeovers
  • 10% Higher overall plant output

Results shown are project targets

Industry questions

Why does a 30% glass-filled PA66 part always show glass read-through on the surface?
Glass read-through usually comes from fibres concentrating at the mould surface while the melt and the wall are at very different temperatures. Fibre loading, screw shear, injection speed and mould temperature all feed into it, so no single parameter fixes it. Questions like this are already covered by trial records and improvement notes in the knowledge base, and the answer pulls up the relevant process cards with sources.
A customer wants PPAP under IATF 16949 — do our TDS and test reports line up?
PPAP asks for drawings, material certificates, dimensional and performance test data and process capability — a whole evidence package, of which the TDS is one small piece, and everything has to match the exact grade and formula revision submitted. Keeping those documents filed by grade and revision means you pull the right version off the checklist instead of sending a customer a superseded TDS.
How do you reach UL 94 V-0 in flame-retardant PC/ABS without losing impact strength?
V-0 comes from a flame retardant that chars the surface or releases radical inhibitors in the flame, and pushing the loading up usually costs impact strength and flow. The practical route is to tune the flame-retardant package together with impact modifiers and compatibilisers rather than adding retardant alone. Formula logic, trial results and failure modes for similar systems can be compared side by side in the knowledge base.
On a highly filled calcium carbonate masterbatch, output is low and strands keep breaking — what do you check first?
Start with feed consistency, then check whether the conveying and mixing sections of the screw are matched, whether the vacuum port is blocked, and what head pressure and strand cooling are doing. Every grade has its own window, and the usual failure is copying the previous grade's process card. Keeping each grade's window and incident history in the knowledge base lets you check before the changeover.
Our formulas are core IP — can this run with no internet connection at all?
Yes. The four core services run on your own servers with Docker Compose or Kubernetes, models can be served locally through Ollama or another internal inference endpoint, and each knowledge base binds its own private OSS or S3 bucket. It works with no route to the internet at all; external sources like Feishu or Notion are only connected when you choose to sync them.
For long-term hot and humid service, how do you choose between PA66-GF30 and PPA?
It depends on whether the hydrolysis exposure is hot water, steam or coolant, and on cost and moulding conditions. Selection records, material comparisons and customer approval status for similar duty are usually already in the knowledge base, so describing the duty clearly gets you the right TDS page and the conclusions from earlier projects.
Parts crack weeks after moulding — is that a drying problem?
It can be. Polar resins like PA degrade hydrolytically when moisture is above spec and the melt is hot, which lowers molecular weight and leaves parts that crack as internal stress relieves. Ejection and assembly stress or a poor compatibiliser choice can do the same. Describing resin, drying conditions, moulding parameters and crack location gets you to the matching failure case far faster than asking 'why do my parts crack'.
How long does loading our existing files take, and do scanned process cards work?
Scanned drawings and the awkward multi-column PDFs common in technical agreements are both handled: OpenDataLoader for complex layouts, PaddleOCR-VL for images and scans, with PDF, Word, Excel, images, EPUB and MHTML all accepted. How long it takes depends on volume and layout quality, so the usual approach is to load one grade family first, check retrieval quality, then roll out.

Other industries