J. Cleeman, R. Malhotra
Keywords: multi-fidelity learning, machine learning, transfer learning, fused filament fabrication
Summary:Modeling the effects of the process parameters on the process or product performance is critical for predictive process design and control. While physics-based modeling has long been the path to such predictions, machine learnt models are of increasing interest recently. But the data-hungry nature of machine learning entails prohibitive experimental cost, is hindered by large oversimplifications of computationally efficiency analytical models, or is limited by the time and cost of creating and executing more accurate numerical models. The situation becomes worse for newer processes or material for which the underlying coupled physics is often not fully known, thus decelerating process or material deployment. Transfer learning has been proposed as a solution for this problem, but the above issues remain since a source model of sufficient fidelity is still needed. This paper takes a first stab at this issue. We propose a physics-informed transfer learning that combines a large amount of data from an extremely low-fidelity but equally computationally efficient process model based on first principles with a small experimental dataset that represents ground truth. We demonstrate this approach by creating a support vector regression model of printed road width in Fused Filament Fabrication. Orders of magnitude reduction in computational cost and 60% reduction in experimental data needed is achieved, while retaining similar or lower error as compared to naïve learning on a much larger experimental dataset. We demonstrate the importance of the number and location of high-fidelity experimental data points and discuss future work in this context.