Engineer design process assisted by explainable deep learning network.

Area of Science:

• Biomechanical Engineering
• Computational Science
• Machine Learning Applications

Background:

• Engineering simulations are crucial for product design but demand significant computing resources.
• Optimizing the use of simulation data is vital for modern industrial product development.
• The mechano-regulatory method, a complex biomechanical theory, is used for dental implant evaluation and post-surgery tissue recovery assessment.

Purpose of the Study:

• To develop a machine learning workflow, primarily using deep neural networks, as an alternative to traditional numerical simulations.
• To apply this workflow to the dental implant design process, specifically evaluating performance and tissue recovery.
• To demonstrate the capability of deep learning networks to replace complex, time-dependent, multi-physical models.

Main Methods:

• A workflow integrating multiple machine learning algorithms, focusing on deep neural networks (DLNs), was proposed.
• The DLN was trained using calibrated simulation data from various conditions, including experimental verification.
• Deep Taylor decomposition was employed for DLN explainability.

Main Results:

• The DLN accurately predicted simulated bone healing history around implants, achieving correlations >0.980 for physical properties and >0.947 for performance indexes.
• Testing AUC values for tissue phenotype classification ranged from 0.90 to 0.99.
• Simulation time was reduced from hours to seconds, and key influencing factors for bone healing were identified.

Conclusions:

• Deep learning networks can effectively replace complex, time-dependent, multi-physical simulation models in engineering design.
• The developed DLN accurately predicts dental implant performance and bone healing dynamics.
• Explainable AI methods revealed critical design features influencing bone healing, offering insights without requiring prior expert knowledge.