[Paper Review] DeepPhysics: a physics aware deep learning framework for real-time simulation

Alban Odot, Ryadh Haferssas+1 more|arXiv (Cornell University)|Sep 17, 2021

Model Reduction and Neural Networks4 references4 citations

TL;DR

This paper presents DeepPhysics, a physics-aware deep learning framework that trains a neural network to predict displacement fields in hyper-elastic materials with micrometer-level accuracy in under a millisecond. By combining a physics-informed loss function, modal analysis-based data generation, and a Hybrid Newton-Raphson algorithm, the method achieves real-time simulation with 54% higher convergence and 20% fewer iterations than standard Newton-Raphson.

ABSTRACT

Real-time simulation of elastic structures is essential in many applications, from computer-guided surgical interventions to interactive design in mechanical engineering. The Finite Element Method is often used as the numerical method of reference for solving the partial differential equations associated with these problems. Yet, deep learning methods have recently shown that they could represent an alternative strategy to solve physics-based problems 1,2,3. In this paper, we propose a solution to simulate hyper-elastic materials using a data-driven approach, where a neural network is trained to learn the non-linear relationship between boundary conditions and the resulting displacement field. We also introduce a method to guarantee the validity of the solution. In total, we present three contributions: an optimized data set generation algorithm based on modal analysis, a physics-informed loss function, and a Hybrid Newton-Raphson algorithm. The method is applied to two benchmarks: a cantilever beam and a propeller. The results show that our network architecture trained with a limited amount of data can predict the displacement field in less than a millisecond. The predictions on various geometries, topologies, mesh resolutions, and boundary conditions are accurate to a few micrometers for non-linear deformations of several centimeters of amplitude.

Motivation & Objective

• To develop a real-time simulation framework for nonlinear elasticity problems that outperforms traditional FEM in speed while maintaining high accuracy.
• To address the limitations of existing methods like domain decomposition and model order reduction by integrating deep learning with physics-based constraints.
• To enable fast, accurate prediction of displacement fields under varying boundary conditions, geometries, and material properties using a data-driven approach.
• To ensure solution validity through a physics-informed loss function and a hybrid optimization strategy that improves convergence.
• To reduce computational cost for real-time applications such as surgical simulation and interactive engineering design.

Proposed method

• Leverages modal analysis to generate an optimized dataset of deformation modes, reducing the number of required training samples.
• Employs a physics-informed loss function that enforces physical consistency, such as boundary conditions and material behavior, during network training.
• Uses a fully connected neural network to learn the nonlinear mapping from boundary conditions to displacement fields.
• Introduces a Hybrid Newton-Raphson algorithm that uses the neural network’s prediction as an initial guess, improving convergence speed and success rate.
• Applies the Newton-Raphson method in a hybrid fashion: if the network prediction is closer to the solution than the first iteration, it is used as the starting point.
• Validates predictions using a residual-based error metric to ensure physical plausibility and solution accuracy.

Experimental results

Research questions

• RQ1 Can a deep learning model trained on a limited dataset predict displacement fields in hyper-elastic materials with micrometer-level accuracy in real time?
• RQ2 How effective is a physics-informed loss function in improving solution validity and generalization across different geometries and boundary conditions?
• RQ3 To what extent does using a neural network prediction as an initial guess improve the convergence rate and iteration count of the Newton-Raphson method?
• RQ4 How does the method generalize to unseen force amplitudes, locations, and geometries beyond the training distribution?
• RQ5 What are the limitations of the approach in handling localized forces or topological changes, and how can they be mitigated?

Key findings

• The neural network predicts displacement fields in less than 1 millisecond, satisfying real-time simulation requirements.
• The method achieves prediction accuracy within a few micrometers for nonlinear deformations of several centimeters in amplitude.
• The Hybrid Newton-Raphson algorithm increases the convergence rate from 46% (standard Newton-Raphson) to 71% across 100 test cases.
• The hybrid approach reduces the average number of iterations from 6.1 to 5.0, with 66% of cases favoring the network prediction as a better initial guess.
• The framework generalizes well to unseen Young’s modulus values, though higher accuracy requires more training samples.
• The method is less effective for highly localized external forces, indicating a need for improved data representation in such cases.