Skip to main content

Design of enlarged phononic bandgap 2.5D acoustic resonator via active learning and non-gradient optimization

Abstract

Identifying the phononic crystal (PnC) with bandgap is a problematic process because all phononic crystals don’t have bandgap. Predicting the Phononic bandgaps (PnBGs) is a computationally expensive task. Here we explore the potential of machine learning (ML) tools to expedite the prediction and maximize the resonator based PnBG. The Gaussian process regression (GPR) model is trained to learn the relationship between complicated shape and band structure of cavity. Bayesian optimization (BO) derives a new shape by leveraging the fast inference of the trained model, which is updated with the augmentation of newly explored structures to escalate the prediction power over performance expansion through active learning. Artificial intelligence (AI) assisted optimization requires a small number of generations to achieve convergence. The obtained results are validated via experimental measurements.

Introduction

An artificially synthesized material with periodically varying acoustic properties is called PnC. Their properties are best described by a phononic band diagram which can have a frequency regime in which excitation is forbidden and called PnBGs. PnBGs are primary characteristics that can be used to engineer novel phononic devices. including filters, cavities, and sensors.

Calculating the PnBG via Finite element method (FEM), plane wave expansion, finite difference time domain method etc., is a computationally expensive task which depends upon numerous parameters and exploring each structural combination with wide parametric range is time consuming and challenging [1, 2].

A computationally inexpensive method is needed to speed up the optimized phononic crystal search process. Now a days the complex relationship between high dimensional output and input can easily be created using ML models, furnishing with enough dataset. In this work we proposed an active learning approach to enhance the prediction tendency and exploring the enlarged bandgap using ML model [3] with smaller dataset as compare to previous studies [4]. For instance, the quick prediction by ML model for enhancing thermal [5], mechanical [6] properties or designing on demand meta structures [7] have been presented successfully. Thus, developing a relationship between the spatial properties and bandgap via training ML model over sufficient large dataset [8] enables us to define the shape variables as a controlling knob to explore enlarged bandgap structures. Here the ML model acts as a surrogate for developing the relationship between cavity structure and bandgap. Our investigation is confined to resonator devices for exploring their potential of suppressing the sound via reflection and absorption, simultaneously. Firstly, we proposed the steps for obtaining the enlarged bandgap structure via active learning. Then comparison with non-gradient optimization is presented. Lastly the experimental realization and verification is discussed [4].

Simulation and experimental method

Data generation

In this study an optimized shaped single opening resonator is searched based on the open curve polygons to achieve the objective of enlarged PnBG. The length of the unit cell is fixed to \({a}_{1}=22 mm\), and the material used for cavity is polylactic acid (PLA). A total of 36 control points is generated inside the unit cell to construct variable thickness cavity which are reduced to 18 control points due to symmetry. Considering n-gon with radius as \(r=cos\left(A\right)/cos\left(B\right),\) where \(A=\uppi /n\) and \(B = \theta_{i} - \frac{2\pi }{n}*\left[ {\frac{{n\theta_{i} + \pi }}{2\pi }} \right]\) To get the maximum possible value of \({r}_{max}\left({\uptheta }_{i}\right)\) we use \({r}_{max}\left({\uptheta }_{i}\right)=r\left({\uptheta }_{i}\right)\cdot {a}_{1}/2sin\left(\uppi /4\right)-{a}_{1}/20\). Thus, the minimum distance of point is \({r}_{min}\left({\uptheta }_{i}\right)={a}_{1}/20\) and maximum distance is given by:

$$r_{max} \left( {\theta_{i} } \right) = \frac{{cos\left( {\frac{\pi }{n}} \right)\frac{{a_{1} }}{2}}}{{cos\left( {\theta_{i} - \frac{2\pi }{n}*\frac{{n\theta_{i} + \pi }}{2\pi }} \right)\sin \left( {\frac{{\uppi }}{4}} \right)}} - \frac{{a_{1} }}{20}$$
(1)

where n is the number of corners i.e., 4\(,{a}_{1}\) is the lattice constant and\({\theta }_{i}\in \left[\text{5,25,45,65,105,120,135}, \text{150,165}\right]\). The structural area can be minimized to circle and maximized to square as shown\*MERGEFORMAT Fig. 1a. Firstly, the shape is prepared in MATLAB then it is transferred to COMSOL for bandgap calculation via livelink. It is noteworthy to mention that with this configuration neck of the structure can also be obtained.

Fig. 1
figure 1

a Shapes of unit cell with minimum to maximum value of radius i.e., \({\text{r}}_{\text{min}}\) to \({\text{r}}_{\text{max}}\). b Workflow chart describing our active learning-based optimization. Here dotted rectangle is referred as single generation

Taguchi method (TM) enables us to generate the 1600 orthogonal combination of structural sample points whose eigenvalues are calculated by FEM using COMSOL having 40 levels. Then GPR model is trained by obtained data and employed to train the relationship between the cavity shape and bandgap. We applied 20 fold cross validation which results in a coefficient of determination (\({R}^{2}\)) equal to 0.96. For global search Genetic algorithm (GA) is applied to check if there is any further improvement possible or not. For predicting enlarged PnBG, BO technique is applied, instead of GA as their search domain is local and structure possessing mean bandgap of data set serve as the initial guess, which gradually shifts towards larger values with the expansion of data set [9]. Thus, the optimal structure is derived using BO aided by expedited inference of GPR algorithm. BO predicts 10 new structures whose ground truth are verified via COMSOL simulation and updates the dataset in order to retrain the ML algorithm. This process repeats five times, which results in the addition of 50 data points in each single iteration. To check the convergence (further improvement) GO is applied (as summarized in Fig. 1b and illustrated in Fig. 2).

Fig. 2
figure 2

Output prediction by ML model at different generations. Each generation possess 50 samples points. First row represents the model check via \({\text{R}}^{2}\) value. Only the results of 2 generations (or 100 points) addition in each successive column are shown

Experimental method

To verify the design experimentally PLA material was used to manufacture the model using 3D printing technique. The sound source is excited by a 100 Hz–12 kHz speaker. The signal is obtained via microphone and fed into the MATLAB program for processing. We can identify the bandgap via applying short time Fourier transform to the signal on complete and fundamental harmonic signal.

Results and discussion

The new candidate’s distribution at the large PnBG range compared to the initial data manifests the efficiency of the proposed method as given in Fig. 3a. There exist some structures having no bandgaps which decreases the performance of ML model as they are not captured fully, which can be seen at the tail end of sorted ranking curve of FEM prediction by ML model (see \* MERGEFORMAT Fig. 3b). In Fig. 3 b this weakness can be clearly seen, but it doesn’t affect since our search in confined to maximize the PnBGs. ML model manages to achieve convergence at \({11}^{th}\) generation, whereas GO algorithm continues its searching process at that point, hence superivority of ML model over direct search GO algorithm is evident from Fig. 4a, b, respectively. Structures possess two types of bandgaps: acoustic and Bragg PnBGs. Here irrespective of type, our objective is to enlarge PnBG size between \(n\) and \(n+1\) bands which can be formulated in terms relative bandgap size as follows:

$$\begin{array}{c}\text{Maximize}:f\left({x}_{BG}\right)=2\frac{\left(\text{min}{\upomega }_{n+1}\left(k\right)-\text{max}{\upomega }_{n}\left(k\right)\right)}{\left(\text{min}{\upomega }_{n+1}\left(k\right)+\text{max}{\upomega }_{n}\left(k\right)\right)}\\ \text{Subject to}:{r}_{min}\left({\uptheta }_{i}\right)\le r\left({\uptheta }_{i}\right)\le {r}_{max}\left({\uptheta }_{i}\right)\end{array}$$
(2)

where \(f\left({x}_{BG}\right)\) denotes the gap-midgap ratio between two consecutive bands, \(n\in \left[1-9\right]\) is the band number,\({\upomega }_{n},{\upomega }_{n+1}\) are eigenfrequencies at wavenumber (k) of target bands.

Fig. 3
figure 3

a Distribution of output at different generations. b The ranking sorted by FEA and ML model

Fig. 4
figure 4

a Maximum bandgap w.r.t to generations using two different approaches. b The band diagram of optimized

To find the bandgap of structure a logarithmic sine sweep technique known as Farina analysis, is used whose logarithmic sinusoidal chirp function is given by:

$$x\left(t\right)=\text{sin}\left(\frac{T{\omega }_{1}}{ln\left(\frac{{\omega }_{2}}{{\omega }_{1}}\right)}.\left({e}^{\frac{t}{T}ln\left(\frac{{\omega }_{2}}{{\omega }_{1}}\right)}-1\right)\right)$$
(3)

where t, T \({\upomega }_{1}\text{and }{\upomega }_{2}\) are the instantaneous time, total time, initial and final angular frequencies respectively. BG can be seen from 7000–10500 Hz with 55db dip occurring at 7900 Hz, which can be seen in complete and 1st harmonic simulation, respectively (see\*MERGEFORMAT Fig. 5a, b). The manufactured 3D printed structure and its model is shown in\*MERGEFORMAT Fig. 5c, d, respectively. By comparing the \*MERGEFORMAT Fig. 5 b, e it is found that numerical simulation and experiment has good agreement. The width of the bandgap is not perfectly same as numerical simulation, which results due to structural imperfection.

Fig. 5
figure 5

a, b Fourier transform of completed and fundamental (\({1}^{\text{st}}\text{ harmonic})\) signal. c, d 3D printed model front view and CAD. e Transmission curve simulation showing the BG in range 7000–10000 Hz

Conclusion

In this work we derived the optimal shape of cavity to enlarge the PnBGs using AI assisted optimization. 1600 different shapes structures were obtained via applying TM. The dimensional parameters were used as predictors for ML model training. BO is combined with ML model to predict the enlarged PnBG structure. ML model is retrained after obtaining 10 points and at the \({50}^{th}\) point the convergence is checked via GA routine. AI assisted resonator structure achieves the larger PnBG value earlier compared to GA. The results are validated using experiments.

Data availability

Data will be made available on request.

References

  1. Tanaka Y, Tomoyasu Y, Tamura SI (2000) Band structure of acoustic waves in phononic lattices: two-dimensional composites with large acoustic mismatch. Phys Rev B—Condens Matter Mater Phys 62:7387–7392. https://doi.org/10.1103/PhysRevB.62.7387

    Article  Google Scholar 

  2. Sigalas MM, Economou EE (1992) Elastic and acoustic wave band structure. J Sound Vib 158:377–382

    Article  Google Scholar 

  3. Ma W, Liu Z, Kudyshev ZA et al (2021) Deep learning for the design of photonic structures. Nat Photonics 15:77–90. https://doi.org/10.1038/s41566-020-0685-y

    Article  Google Scholar 

  4. Lee S, Choi W, Park JW et al (2022) Machine learning-enabled development of high performance gradient-index phononic crystals for energy focusing and harvesting. Nano Energy 103:107846. https://doi.org/10.1016/j.nanoen.2022.107846

    Article  Google Scholar 

  5. Wu S, Kondo Y, Kakimoto M, aki, et al (2019) Machine-learning-assisted discovery of polymers with high thermal conductivity using a molecular design algorithm. npj Comput Mater 5:66. https://doi.org/10.1038/s41524-019-0203-2

    Article  Google Scholar 

  6. Lee S, Zhang Z, Gu GX (2022) Generative machine learning algorithm for lattice structures with superior mechanical properties. Mater horizons 9:952–960. https://doi.org/10.1039/d1mh01792f

    Article  Google Scholar 

  7. Jin Y, He L, Wen Z et al (2022) Intelligent on-demand design of phononic metamaterials. Nanophotonics 11:439–460. https://doi.org/10.1515/nanoph-2021-0639

    Article  Google Scholar 

  8. Zou X, Pan J, Sun Z et al (2021) Machine learning analysis and prediction models of alkaline anion exchange membranes for fuel cells. Energy Environ Sci 14:3965–3975. https://doi.org/10.1039/d1ee01170g

    Article  Google Scholar 

  9. MathWorks Inc. Bayesian Optimization Algorithm. https://www.mathworks.com/help/stats/bayesian-optimization-algorithm.html. Accessed 15 Feb 2024

Download references

Funding

This work supported by a National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIP) (2020R1A2C2009093), and the Korea Environment Industry & Technology Institute (KEITI) through its Ecological Imitation-based Environmental Pollution Management Technology Development Project funded by the Korea Ministry of Environment (MOE) (2019002790007).

Author information

Authors and Affiliations

Authors

Contributions

S. Ibrahim: investigation, Methodology, Formal analysis, Writing–Original Draft. J. Park: project administration, Supervision, Writing–Review & Editing.

Corresponding author

Correspondence to Jungyul Park.

Ethics declarations

Competing interests

J.Park is the editor of Micro and Nano Systems Letters. S.Ibrahim has no conflicts of interest to disclose.

Additional information

Publisher's Note

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

Reprints and permissions

About this article

Check for updates. Verify currency and authenticity via CrossMark

Cite this article

Ibrahim, S.M., Park, J. Design of enlarged phononic bandgap 2.5D acoustic resonator via active learning and non-gradient optimization. Micro and Nano Syst Lett 12, 10 (2024). https://doi.org/10.1186/s40486-024-00202-4

Download citation

  • Received:

  • Accepted:

  • Published:

  • DOI: https://doi.org/10.1186/s40486-024-00202-4

Keywords