Geometry-aware framework for deep energy method:

an application to structural mechanics with

hyperelastic materials
Thi Nguyen Khoa Nguyen1,2,3
with Thibault Dairay1,2, Raphaël Meunier2, Christophe Millet1,3 , Mathilde Mougeot1,4

1 Université Paris-Saclay, Centre Borelli, ENS Paris-Saclay, 91190 Gif-sur-Yvette, France

2 Michelin, Centre de recherche Ladoux, 63118 Cébazat, France
3 CEA, DAM, DIF, 91297 Arpajon, France
4 ENSIIE, 91190 Évry-Courcouronnes, France 1
 1. Context and background
Motivation
• PINNs (Raissi et al. (2019)) have gained much attention in physical modeling:
o Uses a feedforward network as approximator
o Integrates PDEs (strong form) as constraints in the loss function
o Uses automatic differentiation to calculate the derivatives (at colocation points)
• Deep energy method (DEM) (Samaniego et al. (2019), Nguyen-Thanh et al.(2019))
o Use the same principle of PINNs to solve the problems
o Employ weak form of the physical system in the loss function

• However, vanilla PINNs and DEM are capable of inferring the solution on only one
configuration (i.e. fixed IC/BCs, geometry, and other constraints).

Physics-informed neural operators Geometry-aware framework

o Physics-informed DeepOnet (Li et al. (2021)) o PhyGeoNet (Gao et al. (2021))
o Physics-informed Fourier Neural Operator o Physics-informed Point Net (Kashefi et al. (2022))
(Wang et al. (2021)) o Geometry-aware PINNs (Oldenburg et al. (2022))
Inferring on new PDE parameters, or Inferring on new geometries.
new ICs / BCs. 2
 1. Context and background
• Objective: infer the solution on various geometries for structural mechanical
problems.
o An academic test case with reference solution produced by FEniCS.

o A more realistic industrial case with reference solution produced by an in-house solver.

• We propose novel geometry-aware framework for physics-informed models

o Combing weak form of the physical system and shape encoding in the loss function.
o Shape encoding: parametric encoding, PCA, VAE.
o Extension with spatial coordinates or images to represent the geometry for real-world application.

Geometry-aware deep energy method (GADEM) 3

 2. Solid mechanics: physical model
Problem: Find the displacement field 𝒖: 𝛀 → 𝑹𝒅 and a constraint field 𝐏: 𝛀 → 𝑴𝒅 such that:

Strong form Weak form

Potential energy
Constitutive law 𝐏 = 𝜕𝑭 𝑊 𝑭 in 𝛀
(at minimum in the equilibrium state)
Equations of motion Div 𝐏 + 𝐟𝟎 = 𝟎 in 𝛀
Π 𝝓 = 𝑊𝑑𝑉 − 𝒇𝟎 . 𝝓𝑑𝑉 − 𝒇𝟐 . 𝝓𝑑𝐴
𝒖 = 𝒖𝒅 on 𝛀𝐃 𝛀 𝛀 𝛀𝐍
Boundary conditions
𝐏𝐍 = 𝐟𝟐 in 𝛀𝐍 where 𝝓 is a trial function and the displacement 𝒖 fulfills
the boundary condition a priori.

o Very complex systems of PDEs. o Using only first-order differentiations.

o Hard to solve using classical PINNs as involving many high Small computational cost and fast training time.
order differentiations. o Reduce dependencies on PDEs of the base function.

4
 3. Geometry-aware deep energy method
(GADEM)
GADEM is based on the same principle of DEM, and:
o Encodes the geometric knowledge into the model.
o The potential energy of the systems over all geometries are minimized.

Learning Feedforward networks (*) Desired

geometries output

(*) 5 layers, 100 neurons per layer,

5
tanh activation in this work.
 4. Academic experimental design
Linear elasticity problem (small deformation):
o Input: The geometry depends on five parameters
𝑙, 𝑑, 𝑙 , 𝑝 , 𝑝 .
o Configuration: The left side is clamped and the right
side is subjected to a traction 𝑃 = 0, −0.1 𝑁.
Homogeneous, isotropic beams with Young’s
modulus 𝐸 = 1000𝑁/𝑚 , and Poisson’s ratio 𝜈 =
0.3.
o Desired output: The reference
solution with FEniCS by Finite
Element Method (FEM)

6
 4. Academic experimental design
Illustration of GADEM prediction:

Reference
solution
(FEniCS)

GADEM
prediction

Training set Testing set 1 Testing set 2 Testing set 3 Testing set 4
7
 4. Geometric encoding for GADEM: comparison
Performance on training set:

o Parametric encoding: geometric parameters (if

available).
o PCA-Coord: spatial coordinates to represent the
geometries and PCA to encode the geometries.
o VAE-Coord: spatial coordinates to represent the
geometries and VAE to encode the geometries.
o PCA-Image: images to represent the geometries and
PCA to encode the geometries.
o VAE-Image: images to represent the geometries and
Error of the prediction for all geometries in the training set VAE to encode the geometries.
Size of latent vector is fixed .
o GADEM approaches provides good prediction for the solutions (errors vary from 1%-20%)
o The approaches using spatial coordinates performs better than the approaches using
images. 8
 4. Geometric encoding for GADEM: comparison
Performance on testing set
o The approaches using
Test 1 Test 2 spatial coordinates perform
better than the approaches
using images.

o On the same shapes of

geometries as training (test
1 and test 2), VAE-Coord
provides better prediction
Test 3 Test 4
than PCA-Coord.

o On new shapes of
geometries (test 3 and test
4), PCA-Coord provides
better prediction than VAE-
Coord.
9
 5. Toy industrial use case
Tire loading simulation (large deformation with contact)
o We consider the tires (made of rubber, incompressible
material) loading simulation, which follow the Saint-
Venant Kirchhoff hyperelastic model.

o The problem involves contact constraints, which results

additional inequalities.

o As we do not dispose of the geometric

parameters which control the shape of tires,
we consider the following approaches:
 PCA-Coord, VAE-Coord
 PCA-Image, VAE Image
Size of latent vector is fixed .

10
 5. Tire loading simulation
Illustration of GADEM prediction:

Reference
solution
(in-house
solver)

GADEM
prediction

11
Training set Testing set Testing set
 5. Tire loading simulation
Geometric encoding comparison:

Error of the prediction in the training set Error of the prediction in the testing set
o GADEM approaches provide good predictions for the displacements (errors vary from 1-5%)
o VAE-Coord outperforms the others.
o The approaches using spatial coordinates perform better than the approaches using images.
12
 5. Tire loading simulation
Accuracy enhancement with FBOAL
o We apply FBOAL to infer the position of collocation points based
on the potential energy.
o After each 100k epochs, we add and remove 1% of training points.

Density of collocation points for different tires

o FBOAL adds more points near the contact zones.

13
 5. Tire loading simulation
Accuracy enhancement with FBOAL

Error
without
FBOAL

Error with
FBOAL

o The error is significantly reduced with FBOAL, especially at the contact zone. 14
 5. Conclusion and perspectives
o GADEM: a geometry-aware framework for deep energy method.
o GADEM is capable of inferring the solution on various geometries, even on new geometries
that are not included in the training.
o Using spatial coordinates to represent the geometries provides better accuracy than using
images.
o PCA and VAE encoding can provides same performance as parametric encoding.
o FBOAL helps to reduce significantly the prediction errors.

Perspectives:
o Apply GADEM with different IC/BCs.
Related work:
o Nguyen, T.N.K., Dairay, T., Meunier, R., Millet, C. and Mougeot, M., 2023. Fixed-budget online adaptive learning for
physics-informed neural networks. Towards parameterized problem inference. International Conference of
Computational Science.
o Nguyen, T.N.K., Dairay, T., Meunier, R. and Mougeot, M., 2022. Physics-informed neural networks for non-Newtonian
fluid thermo-mechanical problems: An application to rubber calendering process. Engineering Applications of Artificial
Intelligence, 114, p.105176.
15