All Terms

Air trap : a pocket of air the melt cannot push out of the cavity, trapped where two flow fronts close or at a blind rib or boss; shows as a short fill, a burn mark or a weak spot.

APQP (Advanced Product Quality Planning) : the automotive standard that structures a new part's development into five phases, ending in PPAP sign-off.

Backbone : the large pretrained model that produces general-purpose features, reused (often frozen) as the base of a downstream task.

Bayesian optimisation : a sample-efficient search that models the objective probabilistically to choose the next trial; tunes process or design parameters in few experiments.

B-rep (boundary representation) : the native CAD data structure: a solid described by its faces, edges and vertices.

Boss : a raised mounting feature (for a screw or a pin), usually cored out so it does not become a thick section.

CAD (computer-aided design) : software for authoring part geometry (CATIA, NX, Creo, SOLIDWORKS).

CAE (computer-aided engineering) : simulation-based analysis (mould flow, FEA), as opposed to CAD geometry authoring.

Cavity / core : the two halves of a mould: the cavity shapes the outer (visible) surface, the core the inner. The part shrinks onto the core.

Chapter 11 : the US legal process for reorganising an insolvent company while it keeps trading (reorganisation, not liquidation).

Copilot : an LLM assistant embedded inside an engineering tool.

Cycle time : the seconds from one moulded part to the next; cooling usually dominates. The moulder's core economic metric.

Degree of cure (DoC) : how far a thermoset's hardening reaction has progressed; the part can be ejected once a target is reached.

DFM (design for manufacturability) : checking a design against the rules that make it mouldable (draft, wall thickness, rib ratios, undercuts…).

Digital twin : a live simulation model of a physical asset (part, machine, factory) kept in sync with real data, used to monitor and predict its behaviour.

DoE (design of experiments) : running a structured grid of trials or simulations across parameter ranges and picking the best; automated search, not learning.

Draft angle : the slight taper (typically 1–3°) on walls parallel to the mould-opening direction, so the part can be ejected cleanly.

Ejector pins : pins that push the cooled part off the core when the mould opens.

Embedding : a fixed-length vector of numbers representing an item (an image, part or piece of text) so that similar items land near each other in that vector space; the basis for similarity search.

Engineering Intelligence Stack : a five-layer lens (Intent, Knowledge, Reasoning, Execution, Feedback) for locating where intelligence sits, or is missing, in the workflow.

FEA (finite-element analysis) : the standard numerical physics simulation for structures; mould-flow simulation is its plastics cousin.

Fibre orientation : the direction short glass fibres point after moulding; set by the flow, and dominant for the part's stiffness.

Fillet : a smooth rounded transition replacing a sharp internal corner, for better strength and melt flow.

Foundation model : a large model pretrained on broad data (text, images, geometry) then adapted to many tasks, rather than built for one.

Frontier lab : one of the few organisations training the largest, most capable general AI models (OpenAI, Anthropic, Google DeepMind, Mistral…).

Function-calling (tool use) : an LLM emitting a structured request to run a named external function, so it can trigger a search, calculation or action instead of only producing text.

Gate : the small opening where molten plastic enters the part cavity; its position sets the flow pattern, and with it weld lines and packing.

GNN (graph neural network) : an ML architecture operating on graphs; used on B-rep face-adjacency graphs for CAD learning.

Honest-AI test : the single question this study applies to every capability: does it learn from data, or compute from equations?

Hyperscaler : a very large cloud provider (AWS, Microsoft Azure, Google Cloud) running data centres at global scale; the usual host for rented AI models.

KBE (knowledge-based engineering) : rule and template automation inside CAD systems; mature, and honestly not AI.

Labeled data : geometry paired with a recorded outcome (warpage, scrap, tuning history); the scarce ingredient a learned DFM system would need.

LLM (large language model) : a text-trained neural network (GPT, Claude, Mistral…); learns language, not physics.

Machine learning (ML) : models that learn an input→output mapping from data, rather than computing it from fixed equations.

Masterbatch : a concentrated pellet of colour or additive blended into the base polymer at the machine, so moulders buy natural resin plus masterbatch rather than pre-coloured plastic.

Metrology : the measurement of a part's real dimensions (with a coordinate measuring machine, or optical or CT scanning) and their comparison against the CAD model.

Mould-flow simulation : numerically solving how molten plastic fills, packs, cools and warps in a mould; deterministic physics.

Nearest-neighbour search : finding the closest embeddings to a query vector in a large collection; the core lookup operation behind retrieval.

OEM (original equipment manufacturer) : the vehicle maker (BMW, Toyota, VW…), as opposed to its suppliers.

Ollama : a tool for running open-weight LLMs locally on one machine behind an OpenAI-compatible API; the offline counterpart to a hosted model.

Optimization : automated search over a design or parameter space (topology optimization, DoE); it searches, it does not learn.

Overmoulding / encapsulation : moulding polymer around an insert (e.g. a circuit board) to seal and protect it.

PINN (physics-informed neural network) : an ML model whose training loss includes the physics equations; a way to stretch scarce simulation data.

PLM (product lifecycle management) : the enterprise system of record for designs, changes and releases (ENOVIA, Teamcenter, Windchill).

PPAP (Production Part Approval Process) : the formal sign-off that a part and its process are approved for series production; the end of APQP.

Radius / Radii : the size of a rounded corner or fillet.

RAG (retrieval-augmented generation) : an LLM answering from documents retrieved at question time; the mechanism inside most engineering copilots.

Reinforcement learning (RL) : training a model by trial and error against a reward signal rather than labelled examples; it learns a policy of actions.

Re-ranking : a second, more careful scoring pass applied to a short list of candidates from a fast initial search, to sharpen the final ordering.

Retrieval (similarity search) : "find past parts like this one" over an archive; a genuine data capability that needs no outcome labels.

Rib : a thin stiffening wall added to a moulded part; kept to roughly 50–60% of the base wall thickness to avoid sink marks.

ROM (reduced-order model) : a small, fast model (often ML-trained) reproducing the input→output behaviour of a large simulation.

Rules engine : software that checks a design against fixed, human-written rules (the DFM checkers); deterministic, nothing learned.

Runner / sprue : the melt-delivery channels of a mould: the sprue is the entry, the runners distribute melt to the gates.

Self-supervised (learning) : training a model on unlabelled data using tasks it generates for itself (e.g. predicting a masked part of the input), so it learns useful representations without human-labelled examples.

Short shot : a defect where the cavity fails to fill completely before the melt freezes.

Shrinkage : the volume a plastic loses as it cools (several percent); the root cause of sink marks, warpage and the need for packing.

Side action / lifter : an extra moving mould mechanism required to form an undercut; adds tool cost and failure modes.

Sink mark : a shallow surface depression opposite a thick section, caused by local shrinkage; a cosmetic killer on visible parts.

Solver : the numerical engine inside a simulation tool that computes the physics.

SOP (start of production) : the date a part enters series manufacture.

Spike : a short, time-boxed exploratory task to answer one technical question or produce a defensible estimate before committing to a full build.

Steelman : the strongest honest version of a claim (the opposite of a straw man); each scope note states a vendor's steelman before testing it.

Surrogate model : an ML model trained on a batch of simulation results that then predicts outcomes near-instantly without re-running the physics.

Thermoplastic : plastic that melts and re-solidifies reversibly; the material of standard injection moulding.

Thermoset : resin that cures permanently via a one-way chemical reaction; moulded by transfer or compression processes.

Tier-1 supplier : a company selling systems or components directly to vehicle makers (Bosch, Forvia, ZF, Motherson…).

Toolmaker : the specialist (in-house or external) who designs and builds the mould.

Topology optimization : a numerical method that removes material from low-load regions of a design space over repeated FEA runs; the usual technology behind "generative design" labels.

Transfer moulding : a thermoset process where a pre-measured resin charge is pushed from a pot into a closed, heated mould and held until cured.

Undercut : any feature that would trap the part in a two-half mould (side holes, snap hooks, internal lips); each one needs a side action.

Unsupervised (learning) : finding structure in data that carries no outcome labels (clustering, similarity, embeddings); the setting that lets retrieval work without defect labels.

Vector database : a data store built to hold embeddings and answer nearest-neighbour queries efficiently at scale (e.g. FAISS, Milvus, Qdrant).

Virtual sensor : an ML model predicting a quantity that cannot practically be measured directly, from signals that can.

vLLM : a popular open-source engine for serving large language models efficiently at scale; often the software running behind an engineering copilot's chat interface. The "v" is commonly associated with "virtual" (from its virtual-memory-style paged attention), though not officially defined.

Warpage : the part emerging bent or twisted from uneven shrinkage; the hardest defect to predict and the most expensive to fix.

Weld line : the visible seam, and weak zone, where two melt fronts met and knitted imperfectly; placed (via gating) somewhere invisible when possible.

Zero-shot : using a pretrained model on a new task or dataset with no task-specific training examples; contrasted with fine-tuning on that data.