What question did this study set out to answer?

To address the challenge of predicting coating properties across different materials using transfer learning methods.

March 25, 2026Open Access

Transfer Learning for Interpretable Prediction of HVOF Coating Properties

Puntos clave

To address the challenge of predicting coating properties across different materials using transfer learning methods.
Analyzed differences in coating properties using machine learning algorithms.
Implemented transfer learning approaches to adapt existing models to new materials.
Utilized linear residual models for efficient adaptation of predictive models.
Employed symbolic regression models with re-optimizing coefficients.
Linear residual models achieved accurate adaptation of the model for new coating materials.
Maintained model interpretability while significantly reducing experimental effort.
Re-optimizing coefficients in symbolic regression was shown to be effective.

Resumen

High velocity oxygen fuel (HVOF) spraying is used to create protective coating layers on workpieces to achieve specific surface properties. Given the process parameters and a fixed coating material, the resulting coating properties can be accurately predicted using machine learning algorithms. However, since the coating material strongly influences the outcome, it is unclear how well a model trained on one specific coating material can predict properties for a different material. In this work, we demonstrate that significant differences in the marginal distributions between coating materials make direct application of existing models challenging. We present several transfer learning approaches to adapt existing models to new coating materials. Our results show that simple methods, such as linear residual models, are already sufficient to achieve accurate adaptation, drastically reducing the experimental effort for new materials and preserving knowledge from the original model. Furthermore, for symbolic regression models, we show that re-optimizing numeric coefficients is an effective and computationally efficient strategy that also maintains model interpretability.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo