The shear piezoelectric coefficient e15 governs the shear-mode electromechanical coupling in piezoelectric micromachined ultrasonic transducers (PMUTs) and surface acoustic wave (SAW) devices, but measuring it typically requires destructive resonance testing on dedicated specimens. On simulated data, e15 can instead be recovered non-destructively from sparse, noisy subsurface wavefield observations using a constrained physics-informed neural network (PINN). By building exact spatial distance functions and temporal multipliers directly into the network output, the clamped, grounded boundaries and resting initial state hold automatically, leaving the loss with only the interior equations of coupled electro-elastodynamics and the data misfit. Non-dimensional scaling brings all governing coefficients within one decade of O(1), and the setup removes the boundary and initial-condition penalty weights that soft-constraint methods must tune. In a six-way comparison (plane-strain PZT-5H, manufactured wavefield, 30 dB noise), a soft-penalty baseline with adaptive weights converges to an erroneous e15 in four of the five runs (51.7% mean error), and fixed penalties reach an error of 0.85% only after a second-order optimization stage. The constrained formulations reach mean e15 errors between 0.15% and 0.51% with held-out displacement errors below 1%. Within these variants, the spatial multiplier gives rapid first-order convergence, and the quadratic temporal multiplier lowers the early-time electrostatic residual by a factor of 6.8 versus a linear one. On finite-element data along one subsurface line, the full formulation recovers e15 with 1.19% mean error at 30 dB and keeps it between 0.65% and 2.71% across a 40 to 0 dB noise sweep. Finally, training on displacements alone still recovers e15 with mean errors of 0.45% on finite-element and 1.64% on manufactured data.
Figures & tables
Figure 1: Schematic of the two-dimensional plane-strain computational domain for the PZT-5H medium. The perimeter is fully clamped ( ui=0 ) and electrically grounded ( ϕ=0 ), enforcing strict Dirichlet boundary conditions across the entire domain boundary. The physical orientation defines the bottom boundary at x3=0 and the top boundary at x3=1 . For the non-destructive inverse evaluation, acoustic data are taken exclusively from a subsurface observation plane at normalized height x3=0.9 (1 mm below the top surface of the 10 mm domain), with Ns=5,000 space–time observation samples (not physical sensors) in the manufactured-solution tests.
Parameter
Symbol
Value
Density
ρ
7500 kg/m 3
Stiffness (11)
c11E
12.6×1010 Pa
Stiffness (12)
c12E
8.02×1010 Pa
Stiffness (13)
c13E
5.3×1010 Pa
Stiffness (33)
c33E
11.7×1010 Pa
Stiffness (44)
c44E
2.30×1010 Pa
Table 1: Material constants for PZT-5H used in the 2D plane-strain model, adapted from the IEEE piezoelectricity standard [ 15 ] , with c13 set to the value used throughout this work (in the network, the manufactured sources, and the COMSOL model). While c12E=8.02×1010 Pa is a non-zero property of the 6mm crystal class, its corresponding term identically vanishes from the governing equations under the strict plane-strain kinematic constraint ( S2=0 ).
Figure 2: Topology of the constrained multiphysics neural network. The raw network outputs ( Nθ ) are algebraically multiplied by the exact spatio-temporal distance function ( Dbc⋅t2 ). This forward-pass constraint mathematically guarantees the enforcement of the clamped Dirichlet boundaries and a strictly zero-velocity initial state prior to gradient evaluation, completely removing boundary and initial conditions from the loss function.
Figure 3: Trajectories of the recovered e15 for the two soft-penalty baselines on five seeds (30 dB noise, true value 17.0 C/m 2 , dashed line). The horizontal axis shows the Adam steps (0–15,000) followed by the L-BFGS function evaluations. (a) Baseline A with adaptive weights: four runs settle between 25.6 and 31.9 C/m 2 and one reaches 17.3 C/m 2 . (b) Baseline B with all penalties fixed at 5000: the parameter overshoots to between about 59 and 79 C/m 2 , relaxes to 18.3 – 19.0 C/m 2 under Adam, and ends at 17.1 – 17.2 C/m 2 after L-BFGS.
Figure 4: Trajectories of the recovered e15 for the four constrained variants (panels (a)–(d) show C1–C4, five seeds each, 30 dB noise). The dashed line is the true value 17.0 C/m 2 . The horizontal axis shows the Adam steps followed by the L-BFGS function evaluations, and the vertical axis is cut at −15 and 85 C/m 2 . C3, with the spatial factor only, converges smoothly. The variants that contain the factor t2 (C1, C2, and C4) show large excursions before they settle.
Arm
Built into the network
Final error (%)
After Adam (%)
u3 error (%)
Boundary rms
Early-time residual
A
none (adaptive penalties)
51.7±31.3
51.7±31.3
85.8
6.9×10−2
–
B
none (fixed penalties)
0.849±0.245
9.39±1.63
1.50
1.1×10−3
–
C1
Dbct2 on all fields
0.458±0.171
5.14±4.40
0.74
0 (exact)
1.27×10−2
C2
Dbct2 on u1,u3 ; Dbct on ϕ
0.515±0.165
3.90±4.01
0.90
0 (exact)
8.64×10−2
C3
Dbc only
0.163±0.054
0.285±0.115
0.69
0 (exact)
4.62×10−2
C4
t2 only
0.145±0.130
3.42±0.60
0.61
5×10−4
9.1×10−3
Table 2: Controlled comparison of six training formulations on identical data (MMS, 30 dB noise, five seeds). Errors are mean ± sample standard deviation ( N−1 ) of the absolute percentage error in e15 , and “after Adam” is the error before the L-BFGS stage. The u3 error is the mean relative L2 error on 20,000 held-out space–time points. The boundary column is the mean root-mean-square value of uˉ3 on the boundary faces, where the exact value is zero. The early-time residual is the root-mean-square Gauss-law residual for t≤0.05 , evaluated with the true manufactured sources, and is not computed for A and B. A and B are soft-penalty baselines, C1 uses Dbct2 on all fields, C2 uses Dbct on ϕ and Dbct2 on u1 and u3 , C3 uses Dbc only, and C4 uses t2 only.
Figure 5: Parameter discovery using the topologically constrained multiphysics PINN, evaluated against an asymmetric manufactured wavefield subject to 30 dB additive Gaussian noise. (a) The predicted dimensionless displacement field uˉ3 at tˉ=0.5 captures the asymmetric spatial structure arising from the independent wavenumber assignment sin(4πx1)sin(2πx3) , which distributes four antinodal columns across two antinodal rows, with all amplitudes exactly zero at the domain perimeter. (b) A cross-sectional profile at xˉ1=0.125 ( x1=1.25 mm), where uˉ3 has an antinode in x1 , shows the reconstruction of the coupled electromechanical response across the independent spatial frequencies assigned to each field. The dotted curves are the exact manufactured fields. (c) The optimization dynamics show both the internal physics loss ( LPDE ) and the noisy observation data loss ( LData ) minimizing continuously through the sequential Adam and L-BFGS stages. (d) The inverse parameter trajectory for the shear coefficient e15 . The network converges to 16.975 C/m 2 , representing a 0.15% error relative to the true target of 17.0 C/m 2 . The five-run mean error is 0.444%±0.229% .
Figure 6: Parameter discovery using the identical constrained multiphysics PINN applied to COMSOL FEM transient data of the same forced problem with 30 dB additive Gaussian noise. (a) The predicted displacement field u3 at t=5.0μs (normalized tˉ≈0.876 ) reproduces the asymmetric 8-lobe standing wave pattern across the domain and vanishes at the clamped boundaries, as enforced by Dbc⋅t2 . (b) A cross-sectional profile evaluated explicitly at x1=1.25 mm (the primary spatial antinode of the u3 wavefield) shows co-phase behavior of the displacement and the potential. The dotted curves are the exact manufactured fields. The dual axes show the physical amplitudes ( u3 in pm, ϕ in V). (c) Both LPDE and LData decrease through the Adam phase, apart from minibatch noise, and stabilize under L-BFGS. (d) The inverse parameter trajectory for the shear coefficient e15 . The network converges from an initial estimate of 2.33 C/m 2 to 16.873 C/m 2 , representing a 0.75% error relative to the true target of 17.0 C/m 2 .
λ
Mean e15
Mean Absolute Error
(C/m 2 )
(%) ±1σ
1000
17.254
1.497±0.750
2500
17.039
0.952±0.219
5000
16.981
0.444±0.229
10000
16.970
0.175±0.045
Table 3: Sensitivity of the recovered e15 to the data-weighting hyperparameter λ . Results are reported as mean ±1σ over five independent training runs evaluated on the asymmetric MMS under 30 dB additive Gaussian noise. The λ=5000 row is the five-run MMS study of Sec. IV.3 . The true dimensional reference value is e15true=17.0 C/m 2 .
Figure 7: Fields along the observation line x3=9 mm at t=5.0μs for the representative finite-element run. (a), (b) Clean FEM field, every tenth noisy observation, and the PINN prediction for u3 and ϕ . (c) Pointwise difference between the PINN and the clean FEM field, as a percentage of the peak value of each field.
Dataset
Mean e15
Mean Error (%)
(C/m 2 )
±1σ
MMS
16.981
0.444±0.229
FEM, sweep schedule
16.810
1.189±0.651
FEM, long schedule
16.837
0.959±0.309
Table 4: Inverse e15 identification with the full constrained formulation on the manufactured (MMS) and finite-element (FEM) data. Results are reported as mean ±1σ over N=5 independent training runs evaluated at 30 dB additive Gaussian noise. The error column reports the mean of individual absolute percentage errors ⟨∣e15,i−e15true∣/e15true⟩ across the N=5 runs. The true dimensional reference value is e15true=17.0 C/m 2 .
Figure 8: Statistical noise sensitivity of the constrained PINN inverse e15 discovery across nine SNR levels from 40 dB down to 0 dB. The y-axis displays the absolute percentage error in the recovered e15 on a logarithmic scale, and the x-axis is inverted to reflect increasing noise. The solid line represents the mean absolute error evaluated over N=5 training runs per noise level, while the shaded region denotes the ±1σ band, which is truncated at the lower axis limit where the mean minus one standard deviation is not positive.
SNR (dB)
Mean e15 (C/m 2 )
Error (%) ±1σ
40
16.973
0.646±0.280
35
16.838
1.249±0.615
30
16.810
1.189±0.651
25
16.952
0.779±0.532
20
16.861
0.954±0.399
15
16.856
0.847±0.500
Table 5: Statistical noise sensitivity of the constrained PINN inverse e15 discovery across nine SNR levels. Results are reported as mean ±1σ over N=5 independent training runs per noise level, each evaluating a completely reinitialized network against a freshly sampled Gaussian noise realization. The error column reports the mean of individual absolute percentage errors ⟨∣e15,i−e15true∣/e15true⟩ across the N=5 runs. The true reference value is e15true=17.0 C/m 2 .
Figure 9: Trajectories of the recovered e15 for five seeds at 30 dB noise. (a) FEM data with all three fields observed. (b) FEM data with displacement observations only. (c) MMS data with displacement observations only. The horizontal axis shows the Adam steps followed by the L-BFGS function evaluations, and the Adam stage ends at step 15,000. The dashed line is the true value of 17.0 C/m 2 , and the vertical axis is cut at −15 and 85 C/m 2 .
Data
Observed fields
Final error (%)
After Adam (%)
u1
u3
ϕ
MMS
u1,u3,ϕ
0.458±0.171
5.14±4.40
0.99
0.74
0.71
MMS
u1,u3
1.64±0.30
31.5±10.3
1.57
1.15
1.43
FEM
u1,u3,ϕ
0.959±0.309
5.68±3.20
0.96
0.88
0.64
FEM
u1,u3
0.452±0.727
19.0±7.1
1.00
0.88
0.89
Table 6: Recovery of e15 with and without the potential in the data (30 dB, seeds 1 to 5). Errors are mean ± sample standard deviation ( N−1 ) of the absolute percentage error in e15 , and “after Adam” is the error before the L-BFGS stage. The last three columns are the mean relative L2 errors of the fields (%) against the exact wavefield at 20,000 held-out space–time points. The MMS rows with all fields are arm C1 of Table 2 . The FEM rows use the schedule of the representative run.
Physics-Informed Neural Networks present a novel approach in SciML that integrates physical laws in the form of partial differential equations directly into the NN through soft constraints in the loss function. This work studies the application of PINNs to solve a one dimensional coupled electro-elastodynamic system modeling linear piezoelectricity in stress-charge form, governed by elastodynamic and electrodynamic equations. Our simulation employs a feedforward architecture, mapping space-time coordinates to mechanical displacement and electric potential. Our PINN model achieved global relative L2 errors of 2.34 and 4.87 percent for displacement and electric potential respectively. The results validate PINNs as effective mesh free solvers for coupled time-dependent PDE systems, though challenges remain regarding error accumulation and stiffness in coupled eigenvalue systems.
Suhas Suresh Bharadwaj, Reuben Thomas Thovelil
Department of Electrical and Electronics Engineering, BITS Pilani, Dubai Campus, Dubai, United Arab Emirates
Piezoelectric composites are widely used in sensors, actuators, transducers, and energy-harvesting devices because their effective electromechanical performance can be tailored by combining constituent phases and microstructural architecture. However, conventional computational homogenization based on direct numerical simulation (DNS) is computationally expensive, particularly for multiscale simulations and material design tasks that require repeated homogenization analyses. To address this limitation, this work proposes a piezoelectric deep material network (PDMN) to efficiently homogenize two-phase piezoelectric composites. The proposed framework embeds the governing electromechanical homogenization relations directly into the network architecture, yielding a physics-informed, semi-analytical surrogate that explicitly captures the two-way coupling between the mechanical and electrical fields across constituent phases. The network is trained offline on linear electroelastic datasets and, through a fully coupled Newton--Raphson solution with a consistent electromechanical tangent, subsequently used for efficient online prediction under broader constitutive settings, including nonlinear electroelasticity and history-dependent responses. The framework is validated on two-phase composites of polyvinylidene fluoride (PVDF) and lithium niobate (LiNbO3) with reversed phase arrangements under nonlinear electroelastic loading, and on a viscoelastic--piezoelectric composite exhibiting coupled stress relaxation. Numerical examples show that the proposed PDMN achieves high predictive accuracy while reducing the computational cost by more than three orders of magnitude compared with DNS. The proposed framework, therefore, provides an efficient and reliable surrogate for the multiscale analysis and design of piezoelectric composites.
Ting-Ju Wei, Yen-Ming Lu, Chuin-Shan Chen
Department of Civil Engineering National Taiwan University Taipei, Taiwan · Department of Civil Engineering Department of Materials Science and Engineering National Taiwan University Taipei, Taiwan
Physics-informed neural networks (PINNs) provide a promising framework for solving partial differential equations while embedding the underlying physical laws directly into the learning process. This study presents a PINN-based framework for modeling transient elastodynamic wave propagation in bimaterial systems governed by the axisymmetric equations of linear elasticity. A steel-aluminum specimen representative of a Split Hopkinson Pressure Bar configuration is considered, and the governing elastodynamic equations, together with the corresponding initial, boundary, and interface conditions, are incorporated directly into the network through a physics-informed loss function. High-fidelity finite-element simulations performed using ANSYS Workbench Explicit Dynamics are used for validation and as supplementary data constraints during training. The proposed framework accurately predicts wave transmission and reflection across the bimaterial interface and reproduces axial and radial displacement histories, face-averaged responses, and the dominant stress and strain evolution with close agreement to the finite-element solutions. The trained network further demonstrates the ability to predict wave responses at previously unseen time instants and for modified material properties without requiring additional finite-element simulations, providing a continuous surrogate model for elastodynamic analysis. Mesh-sensitivity studies confirm numerical robustness, while additional material combinations demonstrate the generality of the proposed methodology. The results show that integrating physics-informed neural networks with explicit finite-element analysis provides an accurate and computationally efficient framework for elastodynamic wave propagation in heterogeneous solids, offering an effective surrogate modeling approach for high-rate solid mechanics and impact engineering applications.
Sonal Ankush Chibire, Jenn-Terng Gau, Bo Zhang
Department of Mechanical Engineering, Northern Illinois University, 1425 W Lincoln Hwy, DeKalb, 60115, IL, USA.