Publicação
A review on physics-informed machine learning for process-structure-property modeling in additive manufacturing
| Resumo: | This article presents a state-of-the-art review of the emerging field of physics-informed machine learning (PIML) models in additive manufacturing for process-structure-property modeling. Additive manufacturing processes hold immense potential for fabricating intricate and complex geometries across diverse applications and material classes. From a quality assurance standpoint, appropriate modeling of process-structure-property relationships of additive manufacturing processes using either physics-based or machine learning (ML)-based approaches has been a topic of intensive research. As an example, ML of data acquired from in-situ sensors is related to flaw formation, e.g., porosity, cracking, or deformation. In recent years, the computational burden of pure physics-based models, the large data set requirement, and their black-box nature, i.e., the lack of interpretability of ML models, have prompted researchers to turn to PIML models. In PIML models, physical insights of the additive manufacturing process gained from various means are integrated with ML models, resulting in a more robust and interpretable framework for both process and microstructure evolution. A key delineator is the source of physical knowledge to be fused into PIML models, which can be obtained either from governing physical equations, data-centric feature extraction without implementing any physical equations, or a hybrid of the two foregoing. Within this review, we stratify PIML models based on the method used for the fusion of physical knowledge to ML models, into three categories, namely: (i) physics-based feature engineering, (ii) physics-based architecture shaping of ML models, and (iii) physics-based modification of the loss function of the ML models. For each of these categories, we further delineate the source of physical knowledge, ML models, integration approach, and data-set requirement, among others. A comparative analysis of the reviewed studies is presented and critically discussed, while the potential research gaps, along with future research directions on developing PIML models for different AM technologies are outlined. |
|---|---|
| Autores principais: | Faegh, Meysam |
| Outros Autores: | Ghungrad, Suyog; Oliveira, João Pedro; Rao, Prahalada; Haghighi, Azadeh |
| Assunto: | Additive manufacturing Physics-based architecture Physics-based feature engineering Physics-based loss function Physics-informed machine learning Process-structure-property relationships Strategy and Management Management Science and Operations Research Industrial and Manufacturing Engineering |
| Ano: | 2025 |
| País: | Portugal |
| Tipo de documento: | recensão |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | This article presents a state-of-the-art review of the emerging field of physics-informed machine learning (PIML) models in additive manufacturing for process-structure-property modeling. Additive manufacturing processes hold immense potential for fabricating intricate and complex geometries across diverse applications and material classes. From a quality assurance standpoint, appropriate modeling of process-structure-property relationships of additive manufacturing processes using either physics-based or machine learning (ML)-based approaches has been a topic of intensive research. As an example, ML of data acquired from in-situ sensors is related to flaw formation, e.g., porosity, cracking, or deformation. In recent years, the computational burden of pure physics-based models, the large data set requirement, and their black-box nature, i.e., the lack of interpretability of ML models, have prompted researchers to turn to PIML models. In PIML models, physical insights of the additive manufacturing process gained from various means are integrated with ML models, resulting in a more robust and interpretable framework for both process and microstructure evolution. A key delineator is the source of physical knowledge to be fused into PIML models, which can be obtained either from governing physical equations, data-centric feature extraction without implementing any physical equations, or a hybrid of the two foregoing. Within this review, we stratify PIML models based on the method used for the fusion of physical knowledge to ML models, into three categories, namely: (i) physics-based feature engineering, (ii) physics-based architecture shaping of ML models, and (iii) physics-based modification of the loss function of the ML models. For each of these categories, we further delineate the source of physical knowledge, ML models, integration approach, and data-set requirement, among others. A comparative analysis of the reviewed studies is presented and critically discussed, while the potential research gaps, along with future research directions on developing PIML models for different AM technologies are outlined. |
|---|