The purpose of this part of the project is to develop a pre-design tool for pulsating heat pipes. It should be able to predict its performance based on its geometry and its operating conditions, and the result should be available in a few seconds.

Use of machine learning

As full (or even simplified) numerical simulations take several hours to run, the proposed solution is to develop a mathematical macro model. Such a model do not take into account the physics of the PHP. It is trained using machine learning from a database gathered by experiment or numerical simulations. Once trained over known configurations, the model is able to predict the performance of a new one.

Several regression methods have been tested :

  • Polynomial regression
  • Gaussian Process Regressor
  • Artificial neural network
  • Polynomial neural network
  • Kriging
  • Decision Tree

So far, decision trees and gaussian process regressor have shown the best results and predict the performance on the test set with a relative error of 5%.


Once the model is trained, it can be used to optimize the design and/or the operating conditions of a PHP. Once constraints are fixed, the parameters that minimize the thermal resistance are determined.

The method retained for optimization is the new simplicial homology global optimization (SGHO) as it allows to efficiently determine the global minimum for problems in high dimensions.