See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/371947101

A physics-informed machine learning model for inverse reconstruction of

dynamic loads

Conference Paper · April 2023

DOI: 10.2749/istanbul.2023.0315

CITATIONS READS

0 25

3 authors:

Gledson Rodrigo Tondo Igor Kavrakov

Bauhaus-Universität Weimar University of Cambridge
3 PUBLICATIONS 1 CITATION 36 PUBLICATIONS 301 CITATIONS

SEE PROFILE SEE PROFILE

Guido Morgenthal
Bauhaus-Universität Weimar
189 PUBLICATIONS 2,104 CITATIONS

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Dynamic analysis of vessel impact on engineering structures View project

GRK 1462 International Workshop "Coupled Numerical and Experimental Models in Structural Engineering" View project

All content following this page was uploaded by Gledson Rodrigo Tondo on 20 July 2023.

The user has requested enhancement of the downloaded file.

 Istanbul IABSE Symposium (ISTBR 2023)
April 26 – 28, 2023
Istanbul

A physics-informed machine learning model for reconstruction of

dynamic loads
Gledson Rodrigo Tondo
[email protected]
Bauhaus-Universität Weimar
Weimar, Germany

Igor Kavrakov
[email protected]
Bauhaus-Universität Weimar
Weimar, Germany

Guido Morgenthal
[email protected]
Bauhaus-Universität Weimar
Weimar, Germany

ABSTRACT

Long-span bridges are subjected to a multitude of dynamic excitations during their life span.
To account for their effects on the structural system, several load models are used during design to
simulate the conditions that the structure is likely to experience. These models are based on different
simplifying assumptions and are generally guided by parameters that are stochastically identified
from measurement data, making their outputs inherently uncertain. This paper presents a probabilistic
physics-informed machine learning framework based on Gaussian process regression for
reconstructing dynamic forces based on measured deflections, velocities, or accelerations. The model
can work with incomplete and contaminated data and offers a natural regularization approach to
account for noise in the measurement system. An application of the developed framework is given
by an aerodynamic analysis of the Great Belt East Bridge. The aerodynamic response is calculated
numerically based on the quasi-steady model, and the underlying forces are reconstructed using
sparse and noisy measurements. Results indicate a good agreement between the applied and the
predicted dynamic load and can be extended to calculate global responses and the resulting internal
forces. Uses of the developed framework include validation of design models and assumptions, as
well as prognosis of responses to assist in damage detection and structural health monitoring.

Keywords: physics-informed, machine learning, Gaussian process, force reconstruction

1 INTRODUCTION

Assumptions on statistical wind properties, limitations of aerodynamic models and restrictions

in stochastic dynamic analysis are just a few of many sources of uncertainties when modelling
aerodynamic loads during the design phase of a long-span bridge. Furthermore, due to its extended
lifetime and external effects such as climate change, the dynamic forces that are considered during
design are subject to unforeseen changes. These factors motivate the creation of models to reconstruct
dynamic loads based on measurement data.
Several methods for force reconstruction exist in literature, which are generally based on data-
driven techniques [4][5], optimization strategies [10] and defined basis functions [2]. A review of
several of these models is given in [8]. A novel methodology based on stochastic processes is
proposed in this study. The framework combines data-driven models with physics-based
1
formulations, overcomes the necessity of regularization, naturally incorporates measurement noise
properties, and can integrate different data types and sets with different measurement quality.
This study employs a specific aerodynamic model, based on the quasi-steady assumption, to
evaluate the structural response of the Great Belt East Bridge due to an applied wind load. The
evaluated response is contaminated with noise and further used as input to the physics-informed
machine learning model, which in turn yields a stochastic model for the underlying aerodynamic
force. Comparisons and evaluation of the results are provided by comparing the true and
reconstructed signals, and the predictions are coupled with a structural finite element model to
provide insights on structural responses and internal forces due to the aerodynamic loading.

2 GAUSSIAN PROCESS FOR FORCE RECONSTRUCTION

2.1 Physics-informed Gaussian process

The response of a harmonic oscillator to an arbitrary dynamic loading 𝐹 is given by the second-
order inhomogeneous differential equation

𝑚𝑢̈ + 2𝑚𝜁𝜔𝑛 𝑢̇ + 𝑚𝜔𝑛2 𝑢 = 𝐹 (1)

where 𝑚 is the oscillator’s mass, 𝜁 is the damping ratio to critical, 𝜔𝑛 is the oscillator’s circular
natural frequency, and 𝑢 , 𝑢̇ and 𝑢̈ are the displacement, velocity, and acceleration responses,
respectively. These responses are generally directly measurable through sensor devices, and therefore
the displacement response can be modelled as

𝑢 = 𝑓(𝑡) + 𝜀, (2)
2
where 𝑡 is the time and 𝜀 = 𝒩(0, 𝜎𝑛,𝑢 ) the Gaussian noise in the measurement system, characterized
2
by a variance 𝜎𝑛,𝑢 . The same principle of a stochastic process applies to the velocity and acceleration
2 2
signals, where each of them has a particular noise variance 𝜎𝑛,𝑢̇ and 𝜎𝑛,𝑢̈ . Assuming the underlying
deflection response is a stochastic zero-mean Gaussian process, a model can be created such that

𝑢(𝑡) ~ 𝒢𝒫 (0, 𝑘𝑢𝑢 (𝑡, 𝑡 ′ ; 𝜎𝑠 , ℓ) + 𝜎𝑛,𝑢

2
𝛿(𝑡, 𝑡 ′ )), (3)

where 𝛿 is the Kronecker-Delta operator and 𝑘𝑢𝑢 is a covariance kernel parametrized by a standard
deviation amplitude 𝜎𝑠 and a length scale ℓ [7]. Because the displacement response is assumed to be
continuous and smooth in time, 𝑘𝑢𝑢 is herein modelled by the squared exponential (SE) kernel
2
1 𝑡−𝑡 ′
𝑘𝑢𝑢 (𝑡, 𝑡 ′ ; 𝜎𝑠 , ℓ) = 𝜎𝑠2 exp (− 2 ( ) ), (4)
ℓ

which reflects the assumption that similar time indexes should have similar displacement responses.
Since velocities and accelerations are time-derivatives of the displacements and exploiting the fact
and any linear operation to a Gaussian process result in another GP, physics-informed models can
also be created for velocities 𝑘𝑢̇ 𝑢̇ and accelerations 𝑘𝑢̈ 𝑢̈ , as well the respective cross-covariances
between all measurement types. The generated models are used in combination with the measurement
data for training, as shown in Figure 1 (green block).
Combining the derived response kernels with the oscillator model from Eq. (1) gives rise to a
physics-informed cross-covariance model between the dynamic load and the deflection,

𝑑2 𝑑
𝑘𝐹𝑢 (𝑡, 𝑡 ′ ; 𝜎𝑠 , ℓ) = 𝑚 𝑘 𝑢𝑢 + 2𝑚𝜁𝜔𝑛 𝑘 + 𝑚𝜔𝑛2 𝑘𝑢𝑢 , (5)
𝑑𝑡 2 𝑑𝑡 𝑢𝑢

2
while the models for the remaining responses are generated similarly. The force covariance is
calculated by applying Eq. (1) to the second time index of 𝑘𝐹𝑢 , yielding
𝑑2 𝑑
𝑘𝐹𝐹 (𝑡, 𝑡 ′ ; 𝜎𝑠 , ℓ) = 𝑚 2 𝑘𝐹𝑢 + 2𝑚𝜁𝜔𝑛 𝑑𝑡 ′ 𝑘𝐹𝑢 + 𝑚𝜔𝑛2 𝑘𝐹𝑢 . (6)
𝑑𝑡 ′

which can be used for predictions, as seen in Figure 1 (blue and red blocks, respectively).

Figure 1: Framework for the physics-informed Gaussian process: (green) models for measured data
are created and jointly trained. A force model (blue) is built combining the previous models and the
harmonic oscillator’s differential equation, which can be used for predictions (red).

2.2 Optimization of model parameters

Although Gaussian processes are generally regarded as non-parametric models, the choice of
covariance kernel and the noise assumptions lead to free model parameters that shall be identified
based on training data. Assuming measurements from deflections, velocities and accelerations are
available, the parameter set to be identified is defined by 𝜽 = {𝜎𝑠 , ℓ, 𝜎𝑛,𝑢 , 𝜎𝑛,𝑢̇ , 𝜎𝑛,𝑢̈ }, where the
measurement noise can be disregarded if no data from a specific quantity is collected, or extended if
multiple sets of data with different properties are available [9]. Parameter identification is carried
out via maximum likelihood estimation,
1 1 𝑁
𝜽opt = argmax𝜽 log 𝑝(𝒚|𝒕, 𝜽) = argmax𝜽 (− 2 𝒚T 𝑲−𝟏 𝒚 − 2 log|𝑲| − 2 log 2𝜋), (7)

where 𝑁 is the number of data points, 𝒚 = [𝒖, 𝒖̇ , 𝒖̈ ]T collects the training data and 𝑲 is a block-
covariance matrix calculated via the kernel formulations derived in Sec. 2.1,

𝑘𝑢𝑢 (𝒕, 𝒕′ ) + 𝜎𝑛,𝑢

2
𝛿(𝒕, 𝒕′ ) 𝑘𝑢𝑢̇ (𝒕, 𝒕′ ) 𝑘𝑢𝑢̈ (𝒕, 𝒕′ )
2
𝑲=[ 𝑘𝑢̇ 𝑢 (𝒕, 𝒕′ ) 𝑘𝑢̇ 𝑢̇ (𝒕, 𝒕′ ) + 𝜎𝑛,𝑢̇ 𝛿(𝒕, 𝒕′ ) 𝑘𝑢̇ 𝑢̈ (𝒕, 𝒕′ ) ], (8)
2
𝑘𝑢̈ 𝑢 (𝒕, 𝒕′ ) 𝑘𝑢̈ 𝑢̇ (𝒕, 𝒕′ ) 𝑘𝑢̈ 𝑢̈ (𝒕, 𝒕′ ) + 𝜎𝑛,𝑢̈ 𝛿(𝒕, 𝒕′ )

for the specific measurement times 𝒕. Maximization of Eq. (7) is achieved by gradient ascent via
quasi-Newton BFGS optimization. Although not explicitly shown, it is worth noting that the derived
formulation does not require a regular time-step interval, providing flexibility in cases of missing
response time-steps or imperfect sensor synchronization.

2.3 Force prediction

Once the free parameters are identified based on data, the force signal that generated the
measured responses can be reconstructed. The joint distribution of the data collected in 𝒚 and the
force vector 𝑭 is given by

3
 𝟎 𝑲 𝑲∗
𝑝(𝒚, 𝑭) = 𝒩 ([ ] , [ T ]) (9)
𝟎 𝑲∗ 𝑲∗∗

with 𝑲∗ = [𝑘𝐹𝑢 (𝒕∗ , 𝒕′ ), 𝑘𝐹𝑢̇ (𝒕∗ , 𝒕′ ), 𝑘𝐹𝑢̈ (𝒕∗ , 𝒕′ )]T and 𝑲∗∗ = 𝑘𝐹𝐹 (𝒕∗ 𝒕′∗ ), where 𝒕∗ are the time instants
for force prediction. Conditioning the force model on the measurement data yields

𝑝(𝑭|𝒚) = 𝒩(𝝁 = 𝑲T∗ 𝑲−1 𝒚 , 𝚺 = 𝑲∗∗ − 𝑲𝑇∗ 𝑲−1 𝑲∗ ), (10)

which defines the prediction’s mean values 𝝁 and covariance matrix 𝚺 shown in Figure 1.

3 FUNDAMENTAL STUDIES

To evaluate the model’s performance in a controlled and simplified manner, a single-degree-

of-freedom (SDOF) system response subjected to a harmonic force of unit amplitude is numerically
calculated. The oscillator is set with a mass 𝑚 = 1 kg, a damping ratio of 𝜁 = 0.05 and a circular
natural frequency 𝜔𝑛 = 2𝜋 rad/s. The sampling rate of the signal is defined by 𝑓𝑠 = 200 Hz, and the
forcing signal is in resonance with the oscillator. Training data in 𝒚 consist of sparse, regularly-
distributed and noise-free displacement readings, while velocity and acceleration responses are not
provided to the model, leading effectively to only two optimizable parameters 𝜽 = {𝜎𝑠 , ℓ}. The
complete displacement response and the training points (TPs) provided to the model are shown in
Figure 2 (left).

Figure 2: Left: displacement signal and corresponding training points (TPs). Right: true harmonic
force and the model prediction’s mean. The prediction standard deviation are not shown, and tend
to zero as the training data contains no noise.
After the optimal parameters are identified, predictions of the original forcing signal are made
using Eq. (10). The true harmonic force and the probabilistic prediction results are shown in Figure
2 (right). Under idealized conditions, a very good agreement is observed between the true and the
predicted signal, in a mean (𝜇𝐹 (𝑡)) sense. Due to the lack of noise in the training data, represented in
the GP model by setting 𝜎𝑛,𝑢 = 0 m, the standard deviation 𝜎𝐹 (𝑡) of the probabilistic force model
tends to zero, indicating full model confidence in the predictions.

Figure 3: Root mean squared errors of (right) models with different SNRs, for a fixed data set
𝒚 = {𝒖, 𝒖̇ , 𝒖̈ }, and (left) models trained with different data types, for a fixed SNR = 20.

4
 In contrast to the previous assumptions, measurement data is generally contaminated by noise
originating from several possible different sources. The signal-to-noise ratio (SNR) is commonly
employed to quantify the amount of noise in a given signal, herein defined as SNR = 𝐴signal ⁄𝐴noise ,
where 𝐴 is the root mean square function. Moreover, measurement data from all response types is
generally not available, as it is in many cases redundant and not cost-effective. By construction, the
GP model can nevertheless work with missing measured data or incorporate different multi-fidelity
sets of the same data type [9]. To evaluate the effects of both properties, the system is again trained
for data sets containing various SNRs and composed of different combinations of data types.
Although the GP formulation accounts for noise in the measured data via the 𝜎𝑛 parameters,
high-noise signals (low SNRs) can locally modify the force predictions, effectively shifting the mean
prediction in a smoothed region around specific training points, leading ultimately to a high root mean
squared error RMSE. The prediction performance increases rapidly, however, for increasing SNRs,
indicating the influence of sensor quality on the regressed forces, as seen in Figure 3 (left).
Differences between predicted and true force signals also reduce when more than one data type is
available for training, as shown in Figure 3 (right). For cases when measurements contain noise,
providing different data types allows the model to find a balancing point that explains the full
measurement, represented by the covariance kernel in Eq. (8), leading to a stabilized prediction. In
addition, Figure 3 (right) indicates that providing acceleration data for training lead to better
performance when compared to velocity and displacement measurements. While this is the case for
the particular SDOF oscillator considered in this comparison, results are also a function of the system
and applied force properties, and therefore cannot be generalized.

4 AERODYNAMIC FORCES ON THE GREAT BELT EAST BRIDGE

The developed model for force identification is employed to reconstruct the aerodynamic
forces applied to a numerical model of the Great Belt East Bridge in Denmark. The suspension bridge
has a main span of 1624 m and two symmetrical side spans of 535 m each. The deck has a streamlined
cross-section with a chord 𝐵 = 31 m and depth 𝐻 = 4 m. A sketch of the bridge elevation and the
coordinate system for the wind fluctuations and aerodynamic forces on the cross-section level, for
the reduced-order dynamic modelling, are shown in Figure 4, along with the first vertical and
torsional mode shapes.

Figure 4: The Great Belt East Bridge. On top, the elevation (left) and cross-section sketch with the
coordinate system for aerodynamic forces and wind fluctuations (right). At the bottom, the first
vertical (𝑓ℎ = 0.100 Hz, left) and torsional (𝑓𝛼 = 0.278 Hz, right) mode shapes.
For the aerodynamic analysis of the bridge, the wind velocity is separated into a mean
component 𝑈 and fluctuating components 𝑢 and 𝑤 in the horizontal and vertical directions,
respectively. The 2D section is assumed to have three degrees of freedom for horizontal 𝑝, vertical ℎ
and rotational 𝛼 motions. The drag 𝐷, lift 𝐿 and moment 𝑀 forces acting on the deck due to motion,
mean wind and buffeting are obtained using the quasi-steady assumption [1] by

𝐷 = 𝐹𝐿 sin 𝜑𝐷 − 𝐹𝐷 cos 𝜑𝐷 , 𝐿 = 𝐹𝐿 cos 𝜑𝐿 − 𝐹𝐷 sin 𝜑𝐿 , and 𝑀 = 𝐹𝑀 , (11)

with 𝜑𝑖 for 𝑖 = {𝐷, 𝐿, 𝑀} being the dynamic angle of attack, and

5
 1 2 1
2 2 1
𝐹𝐷 = 2 𝜌𝑈𝑟𝐷 𝐵𝐶𝐷 (𝛼𝑒𝐷 ), 𝐹𝐿 = − 2 𝜌𝑈𝑟𝐿 𝐵𝐶𝐿 (𝛼𝑒𝐿 ), and 𝐹𝑀 = 2 𝜌𝑈𝑟𝑀 𝐵 2 𝐶𝑀 (𝛼𝑒𝑀 ) (12)

where 𝜌 is the air density, 𝐶𝑖 is the static wind coefficient and 𝛼𝑒 the effective angle of attack,
calculated by
𝑤+ℎ̇+𝑚𝑖 𝐵𝛼̇
𝛼𝑒𝑖 = 𝛼𝑠 + 𝛼 + 𝜑𝑖 = 𝛼𝑠 + 𝛼 + arctan ( ), (13)
𝑈+𝑢−𝑝̇

where 𝛼𝑠 is the angle of attack at static equilibrium. The resultant velocity 𝑈𝑟𝑖 is calculated as

𝑈𝑟𝑖 = √(𝑈 + 𝑢 − 𝑝̇ )2 + (𝑤 + ℎ̇ + 𝑚𝑖 𝐵𝛼̇ )2, (14)

where 𝑚𝑖 is the aerodynamic centre for 𝑖 = {𝐷, 𝐿, 𝑀} . The dynamic structural response was
calculated for a 10-minute turbulent wind time history. The mean wind speed considered is 30 m/s,
while the isotropic turbulence is defined by an intensity of 10% and generated using the von Kármán
spectrum, with length scales of 200 m for the horizontal direction and 100 m for vertical and
longitudinal directions, respectively. Mechanical admittance in the buffeting response was calculated
using Sears’ model and the aerodynamic centre was defined using Den Hartog’s assumptions. The
static wind coefficients are obtained from [3].

Figure 5: Top: noise-contaminated displacement, velocity, and acceleration responses from the first
vertical mode and corresponding training points (TPs). Bottom: true dynamic force and the
prediction’s mean and 95% confidence interval.
A total of 22 modes are considered in the aerodynamic analysis, and the modal responses are
linearly combined to yield the global coordinate response. Herein, a white noise signal of SNR = 20
is added to each of the modal responses, and the resulting time series is directly used to obtain the
forcing signal. In practice, sensor measurements in global coordinates can be decomposed into modal
components if modal information is available, using techniques such as modal decomposition or
Kalman filtering [6]. A sparse selection of the displacement, velocity and acceleration responses is
used for training, where the data points were selected using a regular time interval of Δ𝑡 = 1.25 s
(cf. Figure 5, top). The resulting prediction of the modal force for the whole 10-minute analysis time,
in the case of the first vertical mode, is shown in Figure 5, bottom.
Similar to the SDOF results, a good agreement is observed between the true modal force and
the GP predictions, even though in the bridge model the force signal contains multiple harmonic

6
components, governed typically by the structural modal frequencies. If a shorter time window is
analysed (cf. Figure 6, left), however, it is evident that the predictions are a smoothened version of
the true force, with the GP effectively acting as a low-pass filter of the original signal and capturing
high-frequency content by its uncertainty measurement. This is also verified by the power spectral
density 𝑆𝐹𝐹 of the force, as shown in Figure 6 (right), where a good agreement is observed until 𝑓 =
2.5 Hz, corresponding approximately to the Nyquist frequency related to the sampling rate of the data
used for training.

Figure 6: Left: detail of the probabilistic prediction in a smaller time window. Right: PSD of the
original and regressed modal force.
Evaluation of the prediction quality for all 22 modes is carried out using the mean squared
error in relation to the normalized true force. Although the modal contributions towards the global
response are functions of the forcing and structural properties, similar RMSE values between 0.04
and 0.10 are observed across all modes, and the predictions have similar quality levels. In practice,
this may not always be true due to the influence of noise in the measurement system.

Figure 7: Comparison between numerical response and probabilistic predictions. Top: deck global
vertical response at midspan. Bottom: internal bending moment response at pylon support.
The force predictions can further be used to evaluate global responses based on the
measurement data, by combining the GP predictions with a finite element model of the Great Belt
East Bridge. Hence, two different examples are now presented. In Figure 7 (top), the global vertical
displacement response at midspan is reconstructed based on the superposition of modal displacements,
in combination with the corresponding mode shapes. In Figure 7 (bottom), the bending moment
values at the pylon support are shown based on the numerical analysis and on the GP predictions.
The model uncertainty is reduced in comparison to the force predictions since the structural response
is generally governed by lower frequencies, which are captured with higher accuracy by the Gaussian
7
 process model, as observed in Figure 6 (right). Furthermore, a good agreement is observed between
the true response and the mean predictions.

5 SUMMARY AND CONCLUSIONS

In summary, a novel stochastic method to reconstruct dynamic forces based on sensor

measurement data has been presented. The model is built based on Gaussian process regression for
machine learning, and relations between physical properties are embedded using well-established
differential equations. This hybrid formulation allows for the use of heterogeneous and multi-fidelity
data during training. Regularization schemes, generally a source of problems in typical optimization
problems, are bypassed by the natural trade-off between data fitting and model complexity provided
by the Gaussian likelihood, which also allows for fully analytical tractability for posterior sampling.
Force predictions were compared and discussed from the perspective of the assumptions taken
for each particular example. The model limitation to represent a wide range of frequency content
correlates with the sampling rate of training data, which may prove problematic when high-frequency
components are expected, as the optimization for GPs scales particularly poorly for an increasing
number of training points. Cases of missing steps or unsynchronized training data can be seamlessly
incorporated by the derived model but were not considered in this study. Moreover, coupling the
probabilistic force predictions with a finite element solver allows for a statistical view of global
responses and, consequently, internal forces. The outputs of this analysis provide valuable insights
into structural performance evaluation and allow for condition diagnostics and prognosis.
In conclusion, the derived framework is a powerful statistical tool for dynamic force
reconstruction and has direct applications in model validation and structural health monitoring.
Although the results presented herein are for a specific case study, it is expected that they extend to
different structural systems and dynamic loading properties.

REFERENCES

[1] Chen, X., Kareem, A. (2001). “Nonlinear response analysis of long-span bridges under turbulent winds”, Journal of
Wind Engineering and Industrial Aerodynamics, 89, 14-15, pp. 1335-1350.
[2] He, W. Y., Wang, Y., Zhu, S., (2018). “Adaptive reconstruction of a dynamic force using multiscale wavelet shape
functions” Shock and Vibration, 2018.
[3] Kavrakov, I., Morgenthal, G., (2018). “A synergistic study of a CFD and semi-analytical models for aeroelastic
analysis of bridges in turbulent wind conditions”. Journal of Fluids and Structures, 82, pp. 59-85.
[4] Liu, Y., Wang, L., Kaixuan, G., Li, M., (2022). “Artificial Neural Network (ANN)-Bayesian Probability Framework
(BPF) based method of dynamic force reconstruction under multi-source uncertainties”, Knowledge-Based Systems,
237, pp. 107796.
[5] Liu, Y., Wang, L., Qiu, Z., & Chen, X. (2021). „A dynamic force reconstruction method based on modified Kalman
filter using acceleration responses under multi-source uncertain samples”, Mechanical Systems and Signal
Processing, 159, pp. 107761.
[6] Nonomura, T., Shibata, H., Takaki, R., (2018). “Dynamic mode decomposition using a Kalman filter for parameter
estimation” AIP Advances, 8, 10, pp.105106.
[7] Rasmussen, C., Williams, C., (2006) Gaussian Processes for Machine Leaning. The MIT Press, Cambridge, MA.
[8] Sanchez, J., Benaroya, H., (2014). “Review of force reconstruction techniques”, Journal of Sound and Vibration,
333, 14, pp. 2999-3018.
[9] Tondo, G., Rau, S., Kavrakov, I., Morgenthal, G., (2022). “Physics-informed Gaussian process model for Euler-
Bernoulli beam elements”, IABSE Symposium: Challenges for Existing and Oncoming Structures, Prague, pp. 445-
452.
[10] Wang, L., Cao, H., Han, X., Liu, J., Xie, Y., (2015). “An efficient conjugate gradient method and application to
dynamic force reconstruction”, Journal of computational science, 8, pp. 101-108.

View publication stats