In this presentation, we detail a novel method for coupling thermal modelling tools (Simcenter 3D Space Systems Thermal) with machine learning (ML) algorithms to develop predictive thermal tools for spacecraft simulations. Additionally we utilize a model reduction method to reduce the complexity of the models which can then be mapped back onto the original high fidelity models. The ML are trained on existing or generated simulation data and used as a predictive tool to drive thermal physics solutions without the need to explicitly resolve the conductance networks over the entire lifetime of the simulation, thus greatly reducing the burden to run explicit simulations.
With some examples, we illustrate that we are able to maintain fidelity of the underlying thermal physics with a number of different ML algorithms for the predictive methods and additionally are able to capture the behavior of the 'real' models by mapping the reduced ML results back on the original
model. This opens the door to facilitate hybrid ML / thermal simulation methods for spacecraft simulations thus reducing the burden of running large models over a large parameter space. We conclude by outlining the pros and cons of this method and our future direction for the hybrid ML / thermal simulation coupling.