MST AI Engineer

How AI reads a P&ID: tags, symbols and lines, and what the engineer still checks

Short answer: a P&ID reader has to recover three kinds of evidence from the drawing: tags (what a component claims to be), symbols (what shape it is) and lines (what connects to what, and in which direction). Tags and symbols can be read fairly directly. Lines are where readers fail, so the useful design is one that flags what it is unsure of and leaves the decision to an engineer.

A P&ID looks like a drawing. To a downstream design tool it is data in disguise: a list of components, the pipes between them, and the direction fluid moves. This article is the first in a series on the MST AI Engineer pipeline. It explains how a P&ID reader works, where it goes wrong, and what we ask an engineer to review before anything moves forward.

Read about P&ID intelligence

Where P&ID recognition sits in the MST AI Engineer pipeline

MST AI Engineer works as four steps: P&ID recognition, then a connection graph, then gas-line-driven 3D part selection from approved PDM parts, then a native, editable SOLIDWORKS assembly. Engineers stay in charge of approvals. The starting scope published on mst-us.ai is semiconductor gas equipment.

This article covers the first step. Everything after it depends on it: a connection that is missing or invented at this stage is carried into every later step.

What a reader has to find

Three kinds of evidence sit on every P&ID, and a reader needs all three.

Tags. A tag is the short alphanumeric label in or next to a symbol, such as PT-101 or PIC-102. Under ANSI/ISA-5.1, the first letter identifies the measured or initiating variable (P for pressure, F for flow, T for temperature, A for analysis), and the letters after it describe the device or function (I for indicate, R for record, T for transmit). So PIC is a pressure indicator-controller and PT is a pressure transmitter. A tag tells you what the component claims to be. It does not tell you where it is or what it connects to.

Symbols. The shape says what kind of thing it is: a manual valve, a check valve, a regulator, a filter, a mass-flow controller. The tag and the symbol must agree. When they do not, something is wrong with the reading or with the drawing.

Lines. Lines carry the topology: which port connects to which, and which way flow goes. For physical design this is the most valuable evidence and the easiest to get wrong.

Evidence on a P&ID, what it tells a reader and how it fails
Evidence	What it tells you	How it fails
Tag (for example PT-101)	What the component claims to be: measured variable and function	Lookalike characters (the letter O and the digit 0), text from a neighbouring symbol, tags in a style the reader has not seen
Symbol	What kind of component it is	Company-specific symbol sheets; the same valve drawn differently by different CAD tools
Line	Which ports connect, and the direction of flow	Crossing read as a junction (or the reverse), broken lines in a scan, an off-page connector with no resolved other end

Tag conventions: ANSI/ISA-5.1, as summarised by ISA and ANSI. MST does not publish a recognition accuracy figure.

Why OCR alone is not enough

Optical character recognition can copy marks off a page. A symbol name with no location, a line with no direction, or a connection with no operating condition is not enough to drive a physical design. That is the idea behind how MST structures the output, described on the MST P&ID intelligence page: every accepted object keeps a pointer to the page and region it came from, candidate identity stays separate from approved engineering identity, and direction and condition stay explicit. The rule is to record observations before permitting inference.

A practical consequence for tags: a tag drawn inside a bubble should be read as part of its symbol, not as loose text nearby, and a reader should prefer the vocabulary the drawing's own legend defines. When two readings are both plausible, the right behaviour is to show both for review, not to pick one.

The hard cases

Most reading problems come from a short list.

Symbol variants. ISA-5.1 defines a standard graphic vocabulary, but companies keep their own symbol sheets, and drawings from different CAD tools draw the same valve differently. A reader trained on one style can stumble on another. The drawing's legend is the first thing to check, which is why a reader should treat the legend as an input and not an afterthought.

Line crossings and junctions. Two lines meeting at a point may be a junction or a crossing. Conventions differ: a hop, a gap, a dot. Research on P&ID digitization treats this as its own problem. The PIDCon dataset (600 P&ID images, 82 component categories and 7,212 component connection pairs) was built for connectivity recognition, and its authors report that position-based methods lose accuracy on intricate line structures such as dashed or jumping lines and on dense component layouts. The PID2Graph dataset represents connection points and line crossings as nodes of their own in its ground-truth graphs. A wrong call creates a connection that does not exist, or removes one that does.

Scan quality. A vector PDF exported from CAD has exact lines and selectable text. A scan has skew, noise, broken lines and blurred characters. The same system will behave differently on the two, and it should say so. Even clean OCR confuses lookalikes such as the letter O and the digit 0; the free MST Drawing to BOM tool warns about exactly this on its page.

Dense layouts and page breaks. Long gas lines wrap across sheets. A connection that leaves one page by an off-page connector must stay unresolved until the other end is found. Guessing here makes a graph that looks complete and is wrong.

What a good reader does when it is not sure

The worst failure is not a missed component. It is a confident wrong answer that nobody looks at. So the useful measure is not how much the machine did, but how many errors reached the next step without being flagged. MST builds around that:

Ambiguity is shown, not hidden. Competing identities or connections are ranked for review and never silently promoted.
Gaps stay open. Cross-page references stay unresolved until both endpoints can be evidenced.
Uncertainty becomes a visible decision. Each result is accepted, sent for review, or refused, and unsupported certainty is not allowed to pass as a result.
Engineers hold authority. Engineers approve the symbol vocabulary, unresolved ambiguity and the release boundaries.

What the engineer reviews

Review is a defined step, not an apology. In order:

Legend check. Confirm what the legend says each symbol means in this project.
Component list. Confirm type and tag for each recognised component, starting with the flagged ones.
Connections and direction. Confirm each pipe's endpoints and the flow direction. Direction is what separates a graph you can design from one you can only look at.
Open items. Resolve or explicitly accept anything left unknown.

What MST is not claiming

MST is not publishing a recognition accuracy figure in this article. A number is only honest when it is measured against labels independently checked by engineers, on drawings the system has not seen, with the failure cases listed beside it. We would rather show the flagged cases than a percentage.

The scope on mst-us.ai is approved semiconductor process-equipment P&ID fixtures, typed symbol vocabularies and project-reviewed rules. Universal recognition of arbitrary drawing conventions is not claimed.

Try the adjacent step today

One neighbouring step is available to use now: the free Drawing to BOM tool on mst-us.ai reads the parts table already printed on a PNG, JPEG or PDF drawing and returns an editable table you can export to Excel or CSV. It reads a table that exists; it does not invent one from geometry, and every row should be checked against the drawing.

Next in this series

How a reader learns a project's symbols: legends, symbol libraries, new symbols, and what happens to an engineer's corrections. After that, the series follows the pipeline through the connection graph, gas-line-driven part selection and the native SOLIDWORKS assembly.

Frequently asked questions

What does a P&ID reader have to extract from a drawing?

Three kinds of evidence: tags (the alphanumeric labels such as PT-101), symbols (the shapes that say what kind of component it is) and lines (which ports connect and which way flow goes). Lines carry the topology and are the easiest to get wrong.

What does a P&ID tag like PIC or PT mean?

Under ANSI/ISA-5.1 the first letter names the measured or initiating variable and the following letters name the device or function. PIC is a pressure indicator-controller and PT is a pressure transmitter.

Why is OCR alone not enough for a P&ID?

OCR copies characters. A tag with no location, a line with no direction or a connection with no operating condition cannot drive a physical design. A P&ID reader also has to keep where each object came from, separate candidate identity from approved identity and keep direction explicit.

What makes P&IDs hard for software to read?

Company-specific symbol sheets, crossings that look like junctions, scan noise and broken lines, dense layouts, and lines that continue on another sheet. Published research on P&ID connectivity reports lower accuracy for position-based methods on dashed or jumping lines and dense layouts.

Does the engineer still review the result?

Yes. Recognition proposes and the engineer confirms: the legend, each component and tag, each connection and its flow direction, and any open items. Anything the reader cannot support should be shown for review, not guessed.

Does MST publish an accuracy figure for P&ID recognition?

No. A figure is only meaningful when measured against engineer-checked labels on drawings the system has not seen, with failure cases listed.

See how MST AI Engineer works

MST AI Engineer takes a P&ID gas-line diagram plus approved PDM parts and equipment rules and builds a native, editable SOLIDWORKS assembly, starting with semiconductor gas equipment. Read the P&ID intelligence page on mst-us.ai.