Rapid AI-based generation of coverage paths for dispensing applications

TL;DR

The paper addresses the bottleneck in TIM coverage path planning by proposing an AI-based end-to-end design-generation approach that infers dispense paths directly from a target cooling area. It introduces a two-part architecture: a pretrained quality model ANN (2D discretization, flow behavior, void detection) and a process model ANN that outputs a fixed 12-parameter path representing the $x$ and $y$ coordinates of a polygonal chain with five segments and six path points, with a feedrate $f = V/l$ to match TIM volume $V$ along length $l$. Training relies on the quality-model objective and unlabeled target-area outlines, enabling real-time inference after decoupling the process model from the quality model. Results demonstrate feasible, air-void-free paths with sub-second inference times, and a clear efficiency advantage over TIMtrace, which can require up to a week of computation while delivering comparable coverage; the approach also hints at potential transfer to other manufacturing processes where inline process-parameter optimization is beneficial.

Abstract

Coverage Path Planning of Thermal Interface Materials (TIM) plays a crucial role in the design of power electronics and electronic control units. Up to now, this is done manually by experts or by using optimization approaches with a high computational effort. We propose a novel AI-based approach to generate dispense paths for TIM and similar dispensing applications. It is a drop-in replacement for optimization-based approaches. An Artificial Neural Network (ANN) receives the target cooling area as input and directly outputs the dispense path. Our proposed setup does not require labels and we show its feasibility on multiple target areas. The resulting dispense paths can be directly transferred to automated manufacturing equipment and do not exhibit air entrapments. The approach of using an ANN to predict process parameters for a desired target state in real-time could potentially be transferred to other manufacturing processes.

Figures (2)

• Figure 1: Approach of our newly proposed training setup. No labeled training data is needed. The process model ANN on the left is trained (indicated by the fire icon) with the help of the quality model ANN on the center-right. The quality model ANN itself is composed of multiple individual models - namely the 2D discretization ANN, the flow behavior ANN and the void detection ANN. During the training of the process model ANN, the weights of the pretrained quality model ANN are frozen (indicated by the ice icon). After training, the process model ANN can predict dispense paths for any given target cooling area.
• Figure 2: Output of the trained AI model for unseen target cooling areas. The upper row shows results for living room shapes as used during training. The lower row shows results for actual product geometries. Green: Target area, grey: compressed state of TIM, light-yellow: dispense path. The percentage of covered target area $C$ is indicated for each area.