Draper Labs Presentation Summary:  Tom Marinis gave a talk entitled “Measurements of Anisotropic Thermal Conductivity of a Multilayer Ceramic Substrate Over a Uniform Spatial Distribution” that was originally presented at IMAPS 2025 in San Diego. This investigation looks at the potential for improving thermal conductivity of multilayer ceramic substrates to support high power processors and sensor systems.  An analytical model was presented that quantifies the increase in thermal conductivity of a substrate that can be realized by aligning the signal traces along the flow of heat rather than normal to it. The impact of conductor orientation on substrate heat conduction was demonstrated by using a wire saw to dice a substrate into 5-millimeter square tiles and measuring their thermal conductivity along three orthogonal directions. 

An approach was described for utilizing machine learning to enhance thermal design as part of a substrate layout process. Finite element analysis was used to compute the thermal conductivity of a collection of 100, 5-millimeter square tiles with different conductor line patterns. Eight quantitative attributes of the patterns such as total conductor length, average conductor width and spacing, number of jogs, terminations, etc. were used to train a three-layer neural network to predict thermal conductivity based on the conductor pattern. The machine learning program was adapted from one published by J. McCaffrey that utilized the Python based PyTorch library.  The thermal performance of a new substrate design could then be quickly estimated from the artwork by dividing it into a tiled array with different machine predicted anisotropic thermal conductivities rather than performing a lengthy finite element analysis.

Draper Labs Slide Deck:
https://docs.google.com/videos/d/1Vdu1AYA6cGnGa86d8cJGmUGswo9ACEjkB5DZPjVcfUo/edit?usp=sharing

Draper Published Paper:
https://docs.google.com/document/d/1vofhAJc6Tx1e_IqC77CbJaUCaptZd7Pf/edit?usp=sharing&ouid=113445902864075247293&rtpof=true&sd=true