Neural Operator and Physics-Informed Deep Learning Approaches for Inverse Design of Composites and Manufacturing Processes

Author

Minglei LuFollow

Date of Award

12-2024

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Mechanical Engineering

Committee Chair/Advisor

Zhen Li

Committee Member

Gang Li

Committee Member

Phanindra Tallapragada

Committee Member

Zhaoxu Meng

Abstract

In this dissertation, artificial intelligence (AI) models are designed and used to accelerate inverse design of composites and manufacturing processes. The critical bottlenecks in machine learning (ML) including data availability, data quality, model generalization and adaptation, interpretability, physical consistency, and the ’black box’ nature of models for the inverse design are addressed. And the proposed AI models are tested under different engineering scenarios. Firstly, a fast deep neural operator (DNO) structure was developed to significantly reduce training time. This model was tested in the context of additive manufacturing, a transformative industrial technology that allows for the creation of materials with complex three-dimensional structures directly from computeraided design models. By constructing interpenetrating phase composites (IPCs) with superior mechanical properties, the DNO model learned the transient response of IPCs under dynamic loading, achieving a 100x speedup over conventional deep operator networks. After training, the DNO predicted transient mechanical responses with over 98% accuracy within one second, significantly expediting IPC structural design. Secondly, a composite neural operator model is developed to unifies the analysis of nonlinear bubble dynamics across macroscopic and microscopic scales. The formation and growth of bubbles during composite material production can significantly impact the material’s integrity and mechanical properties. The model integrates a many-body dissipative particle dynamics (mDPD) approach with a continuum-based Rayleigh-Plesset (RP) model using a novel neural network architecture. By combining a deep operator network to capture the mean behavior of bubble growth with a long short-term memory network to model microscale fluctuations, the composite neural operator achieves high predictive accuracy while effectively capturing size-dependent stochastic behavior. Thirdly, a novel deep learning framework, the enhanced deep neural operator (DNO+), is introduced. The DNO+ model extends the DNO framework by incorporating diverse material properties and process parameters into the predictive model. The model can capture the intricate relationships between system input, material properties, process parameters and the system responses. It is validated in the biomass comminution process. The DNO+ model is used to predict the particle size distribution (PSD) of comminuted biomass in a large knife mill, and achieve great prediction accuracy. Finally, PIDNO+ is developed to integrate physics into the model, taking the advantages of both deep learning model and physics model, so that it can reduce the data requirements and improve physical consistency of the model. Additionally, symbolic regression is employed to enhance model’s explainability, making the ’black box’ nature of the model more transparent and easier for engineers to use. The results demonstrated remarkable accuracy in both calibration and training, providing effective guidance for general comminution design. Furthermore, inverse design of the comminution process is applied. By specifying the desired PSD of the final product, the inverse design method can be used to determine the optimal combination of material properties and processing parameters, ensuring that the manufacturing process is tailored to meet specific performance criteria efficiently and effectively. The approach can also be extended to other engineering fields, providing a powerful framework for the accelerated and accurate design of composites and manufacturing processes with desired properties.

Recommended Citation

Lu, Minglei, "Neural Operator and Physics-Informed Deep Learning Approaches for Inverse Design of Composites and Manufacturing Processes" (2024). All Dissertations. 3810.
https://open.clemson.edu/all_dissertations/3810