Machine Learning (ML) is of increasing interest for modeling parametric
effects in manufacturing processes. But this approach is limited to established
processes for which a deep physics-based understanding has been developed over
time, since state-of-the-art approaches focus on reducing the experimental
and/or computational costs of generating the training data but ignore the
inherent and significant cost of developing qualitatively accurate
physics-based models for new processes . This paper proposes a transfer
learning based approach to address this issue, in which a ML model is trained
on a large amount of computationally inexpensive data from a physics-based
process model (source) and then fine-tuned on a smaller amount of costly
experimental data (target). The novelty lies in pushing the boundaries of the
qualitative accuracy demanded of the source model, which is assumed to be high
in the literature, and is the root of the high model development cost. Our
approach is evaluated for modeling the printed line width in Fused Filament
Fabrication. Despite extreme functional and quantitative inaccuracies in the
source our approach reduces the model development cost by years, experimental
cost by 56-76%, computational cost by orders of magnitude, and prediction error
by 16-24%.

本研究提出了一种基于迁移学习的方法，通过使用大量来源于基于物理的流程模型的计算廉价的数据进行 ML 模型训练，然后在少量代价昂贵的实验数据上进行微调，从而降低了新工艺过程物理模型开发的高成本，并且该方法在 Fused Filament Fabrication 的印刷线宽建模中表现良好，减少了模型开发时间、实验成本、计算成本和预测误差。