01　What makes prototyping difficult? — A CAD drawing and a printed device do not always match

Imagine changing the width of a Y-shaped channel. You revise the CAD model, print a new device, and inspect its dimensions under a microscope. When the result differs from the target, you repeat the cycle.

Designed width: 1,000 µm. Measured after printing: 1,095.5 µm.

The figure on the right compares CAD models with fabricated Y-junction devices. The target overall channel-track width was 1,000 µm (1 mm). The mean printed width was 1,095.5 µm — approximately 9.6% wider.

Overall Y-junction track width, before correction

1,000 µmdesign →1,095.5 µmmeasured

Paper Table II. These values refer to the external channel-track width, not the internal flow-channel width.

Can we identify where the deviation occurs and use that information to improve the next prototype?

02　Our approach — ChatGPT writes the code; Python generates repeatable CAD files

We give the language model the channel type and target dimensions, then use it to develop Python code. After checking and refining the code, we keep the finalized script and adjust numerical parameters such as width and angle to regenerate the design.

Why save and reuse the finalized Python script?

Identical prompts to ChatGPT can generate different code. A checked and finalized Python script, executed under the same conditions, generates the same design files. The LLM helps write the program; the fixed program reproduces the CAD geometry.

A hash is a file’s digital fingerprint. For each geometry, three executions of the finalized script yielded STL files with matching MD5 hashes. This confirmed that the design files were identical across repeated runs.

03　How much faster? — Regenerate a design in seconds

Once the script is ready, changes to channel width, curvature, junction angle, or spiral loop count can be entered as parameters. The updated CAD files are exported within seconds.

Paper Table I; mean of three runs per geometry. Python times cover execution of the finalized script through file export. Manual CAD times cover drafting through STL export.

Developing the initial script took approximately 45–60 minutes of interactive work with ChatGPT. Once created, that script eliminates the need to redraw every variation from scratch, making it especially useful when comparing many design conditions or repeating fabrication cycles.

04　Did the printed channels meet the target dimensions? — Measure, then correct

A fast CAD model is useful only when the manufactured device can be evaluated against its target. We verified the digital geometry, fabricated channels by fused deposition modelling (FDM), and measured the printed devices using digital microscopy.

Measurements separate geometry generation from printing errors

We first checked the STL dimensions in MeshLab, then measured the fabricated channel-track dimensions under a digital microscope. This comparison showed that the dimensional deviations arose primarily during FDM fabrication.

Using measurements to update the next CAD design reduced width deviations

Paper Table IV. These values compare the first and second compensated iterations. Devices were fabricated using PLA on a Bambu Lab A1 printer.

Flow testing showed no visible leakage up to 10 µL/min. Leakage began at 15 µL/min for the spiral mixer, 30 µL/min for the Y-junction, and 35 µL/min for the T-junction. The study tested not only whether the geometry could be generated, but also whether the printed channels carried fluid.

The outcome is a repeatable prototyping workflow: generate CAD, fabricate a device, measure the result, and use that measurement to improve the next design.