Desktop Workbench for AI Analysis of Semiconductor Metrology Images

• Two AI tasks on semiconductor metrology images, in one program
• First, generating a measurement image from a design layout
• Second, judging whether a captured measurement image is normal or abnormal
• Each task split into inference, data preparation and training, so model users and model builders work in the same program
• Operators who never touch the models themselves run a job by picking paths and pressing a button

• Analysis models shipped as command-line scripts, runnable only by someone who knows the argument names and output format
• Measurement info files and image folders laid out in three different structures depending on the site
• Differing folder depth and numbering rules, so one fixed way of reading an image list would not do
• Normal/abnormal labels for training applied by a person looking at each image
• Models still changing, so the screens had to survive a model being swapped out

• Top-level tabs for the two tasks; inference, data preparation, training and browsing as sub-tabs
• Analysis scripts called with a fixed argument contract, their console output mirrored straight into an on-screen log
• Jobs that report progress parsed for that figure and reflected in a progress bar
• Selecting a result row opens that image alongside it, so score and picture are read together
• Missing paths named explicitly when a job is started without them
• Errors shown verbatim, so the point of failure stays visible

Generating a measurement image from a layout

• Open layout → mark the region to cut → choose preprocessing → generate → save
• Region set either by width/height entry or by dragging on the image
• Edge extraction, greyscale and horizontal/vertical flips toggled with the input updating immediately
• Edge extraction tuned through two threshold values

Screening measurement images

• Point at an image folder and a measurement info file → run → per-image scores and verdicts in a table
• Selected image panned and zoomed in the pane beside the table
• Results written as a comma-separated table file, openable in other tools

Preparing data and training

• Distortions applied to source images to build a training set → original and result compared side by side
• Training options edited on screen, then training started with the log streaming live
• An earlier result file nominated as a starting point to resume training

Reading scattered folder structures

• All three layouts of measurement info file and images read into a single list
• Sub-folders walked to collect images, with the original folder structure kept visible in the tree
• Tidied data exported back into whichever of the three structures is wanted

Aligning two images before cutting

• Layout and measurement image overlaid, each with its own size and opacity, and aligned by eye
• Either image moved independently while overlaid, so images at different magnifications can be matched
• Once aligned, the same region cut and saved as three files: layout, measurement, and the overlay

Labelling and comparing

• Labels picked directly in the list, with total, normal, abnormal and unlabelled counts always on screen
• Labels loaded from file remembered even for images not currently open → applied as soon as that folder is opened
• Browser split into two panes for comparing different images, with pan and zoom applied to both at once
• A drag that would not release when the mouse left the image frame fixed down in the display layer
• Brightness distribution, the coordinates of the visible region, image size and file size shown per image

• Analysis runnable by picking paths and pressing a button, without being the person who built the model
• Verdict and image read on one screen → no judgement made from the table alone
• Labelling, training-set generation, training and result review joined inside one program
• Screens unchanged when an analysis script is replaced → the way of working survives a model change
• Site-specific folder structures read the same way, and rewritten into whichever structure is needed
• Open-source licences for the software used readable from inside the program