on Machine Learn., Proc. The principal difference is that oscillations in distant stars cannot be resolved. Additional different variants of VAE are utilized to find a more suitable design that would i) provide reconstructions focused on the sunquake rather than the AR of the input sample, and ii) is able to capture small variations around or inside the AR where sunquakes usually occur. To mitigate the shuffle bias described earlier, we group the input data by event and only perform shuffling after dividing the groups between training and validation data. On average, results of this approach are similar to those of the autoencoder methods. Lin, T.-Y., Goyal, P., Girshick, R., He, K., Dollr, P.: 2017, Focal loss for dense object detection. We note that all of the recovered sunquakes in SC24 maintain roughly the same characteristics. Open Sour. 4)Studies of sunquakes, or helioseismology, have revealed that A)"sunquakes" are caused by similar processes that create earthquakes on the earth. Local helioseismologya term coined by Charles Lindsey, Doug Braun and Stuart Jefferies in 1993[28]employs several different analysis methods to make inferences from the observational data.[2]. Solar Physics The outputs of these models are combined with OD model outputs to derive both temporal and location information of sunquakes for additional datasets. What has studies of sunquakes or helioseismology revealed? In this section we provide the architecture and parameters of two CL models and analyze their predictions. Experiments performed on the holography data indicated that the main challenges in sunquake classification include class imbalance, low-data regime problems, and the inability of autoencoder-based approaches to capture the relevant sunquake features in the latent distribution. (eds.) If the autoencoder objective is extended to include and simultaneously optimize both reconstruction and latent-space distribution losses, the latent state for an observation is closer to others but deviates when necessary to describe distinctive features in the input (Kingma and Welling, 2019). Sunquakes: Probing the Interior of the Sun: Zirker, J. B The outward pressure due to super heated gas. Google Scholar. {\displaystyle \omega ^{2}=gk_{\rm {h}}} To interpret the clustering of the encoded data points, dimensionality reduction techniques are applied, specifically principal component analysis (PCA: Makiewicz and Ratajczak, 1993) and uniform manifold approximation and projection (UMAP: McInnes etal., 2018). This particularly powerful sunquake's ripples emanated from a rather strong solar flare. Granular flows at the solar surface are mostly horizontal, from the centres of the rising granules to the narrow downdrafts between them. Adv. Finally, we apply the random erase transform as described above, generating up to eight gray erasing rectangles of varying proportions. Y These are described in more detail in AppendixA. J. The method is applied to photospheric Dopplergram maps such as those provided by the Michelson Doppler Imager (MDI: Scherrer etal., 1995) onboard the Solar and Heliospheric Observatory (SOHO: Domingo, Fleck, and Poland, 1995) and the Helioseismic and Magnetic Imager (HMI: Schou etal., 2012) onboard the Solar Dynamics Observatory (SDO: Pesnell, Thompson, and Chamberlin, 2012). {\displaystyle R/\ell } In: Lee, D., Sugiyama, M., Luxburg, U., Guyon, I., Garnett, R. Solar Phys. For this ML-oriented work, where absolute intensity measurements tend to lose significance, the most important applicable constraint is the temporal-selection criteria. B) "sunquakes" are caused by similar processes that create earthquakes on Earth. All authors contributed to writing the Abstract and Sections 1 and 5. J. In: Berger-Wolf, T., Chawla, N. To alleviate this, techniques that instruct the model to distinguish between false and correct sunquake signatures need to be employed. B The egression power \(\mathrm{P}(\boldsymbol{r},t)=|\mathrm{H}_{+}(\boldsymbol{r},t)|^{2}\) is extensively used in detecting or studying acoustic sources and absorbers (Ionescu, 2010). Before diving into the ML-driven methodology, we will first describe the processes that data undergo in preparation for model ingestion. A two-step approach combining self-supervised with supervised CL models was found to be robust in capturing sunquake features, as shown in Section3.1. To explain this, we prepared the model in Section4.1.1, which presents none of the above biases, to offer a clean overview of baseline capabilities. Mumford, S.J., Freij, N., Stansby, D., Christe, S., Ireland, J., Mayer, F., Shih, A.Y., Hughitt, V.K., Ryan, D.F., Liedtke, S., Hayes, L., Prez-Surez, D., I., V.K., Chakraborty, P., Inglis, A., Barnes, W., Pattnaik, P., Sipcz, B., Sharma, R., Leonard, A., Hewett, R., Hamilton, A., Manhas, A., MacBride, C., Panda, A., Earnshaw, M., Choudhary, N., Kumar, A., Singh, R., Chanda, P., Haque, M.A., Kirk, M.S., Konge, S., Mueller, M., Srivastava, R., Jain, Y., Bennett, S., Baruah, A., Arbolante, Q., Charlton, M., Maloney, S., Mishra, S., Paul, J.A., Chorley, N., Chouhan, A., Himanshu, Zivadinovic, L., Modi, S., Verma, A., Mason, J.P., Sharma, Y., Naman9639, Bobra, M.G., Manley, L., Rozo, J.I.C., Ivashkiv, K., Chatterjee, A., von Forstner, J.F., Stern, K.A., Bazn, J., Jain, S., Evans, J., Ghosh, S., Malocha, M., Visscher, R.D., Staczak, D., Singh, R.R., SophieLemos, Verma, S., Airmansmith97, Buddhika, D., Alam, A., Pathak, H., Sharma, S., Agrawal, A., Rideout, J.R., Park, J., Bates, M., Mishra, P., Gieseler, J., Shukla, D., Taylor, G., Dacie, S., Dubey, S., Jacob, Cetusic, G., Reiter, G., Sharma, D., Inchaurrandieta, M., Goel, D., Bray, E.M., Meszaros, T., Sidhu, S., Russell, W., Surve, R., Parkhi, U., Zahniy, S., Eigenbrot, A., Robitaille, T., Pandey, A., Price-Whelan, A., J, A., Chicrala, A., Ankit, Guennou, C., DAvella, D., Williams, D., Verma, D., Ballew, J., Murphy, N., Lodha, P., Bose, A., Augspurger, T., Krishan, Y., honey, neerajkulk, Ranjan, K., Hill, A., Kekek, D., Altunian, N., Bhope, A., Singaravelan, K., Kothari, Y., Molina, C., Agrawal, K., mridulpandey, Nomiya, Y., Streicher, O., Wiedemann, B.M., Mampaey, B., Agarwal, S., Gomillion, R., Gaba, A.S., Letts, J., Habib, I., Dover, F.M., Tollerud, E., Arias, E., Briseno, D.G., Bard, C., Srikanth, S., Stone, B., Jain, S., Kustov, A., Smith, A., Sinha, A., Tang, A., Kannojia, S., Mehrotra, A., Yadav, T., Paul, T., Wilkinson, T.D., Caswell, T.A., Braccia, T., yasintoda, Pereira, T.M.D., Gates, T., platipo, Dang, T.K., W, A., Bankar, V., Kaszynski, A., Wilson, A., Bahuleyan, A., Stevens, A.L., B, A., Shahdadpuri, N., Dedhia, M., Mendero, M., Cheung, M., Mangaonkar, M., Schoentgen, M., Lyes, M.M., Agrawal, Y., resakra, Ghosh, K., Hiware, K., Gyenge, N.G., Chaudhari, K., Krishna, K., Buitrago-Casas, J.C., Qing, J., Mekala, R.R., Wimbish, J., Calixto, J., Das, R., Mishra, R., Sharma, R., Babuschkin, I., Mathur, H., Kumar, G., Verstringe, F., Attie, R., Murray, S.A.: 2022, SunPy, Zenodo. B 2020 - 2024 www.quesba.com | All rights reserved. Following Kingma and Welling (2014), we modify our model into a variational autoencoder (VAE), by updating the encoder to also output a distribution in the latent space with explicitly modeled variance. Do they have to give members warning before they bar you? An average coverage of correctly localized sunquake signatures in 44.2% of the total positive frames for singular events. Xu, F., Cervone, G., Franch, G., Salvador, M.: 2020, Multiple geometry atmospheric correction for image spectroscopy using deep learning. To lower the impact of the class imbalance, we experiment with different transforms for augmenting our input data, including: center and random crops, sharpness, color jitter, posterize, invert, auto contrast, solarization, Gaussian blur, normalization, vertical/horizontal flips, and random rotations of 90, 180, 270. {\displaystyle c^{2}=\gamma _{1}p/\rho } ) Quiz October 2014, questions and answers - Astronomy - Studocu Although still unreliable for classification, the reconstructions may prove potentially useful for tasks related to masking the AR with the purpose of reducing noise present in the dataset. Helioseismology Step-by-step explanation "Sunquakes" are studied to determine the interior properties of the Sun much like S and P waves are used to study the interior structure of the Earth. (eds.) B) plumes of hot mantle rising in a hot spot within a plate. Towards the end of the decade, observations also began to show that the oscillation mode frequencies vary with the Sun's magnetic activity cycle.[59]. This event occurs very close to the solar limb, so projection effects are non-negligible in both the AIA emission and in the data used for acoustic-signature identification. The encoding produced by this model is used to perform the classification. Figure10 shows the clustering of embeddings, colored by prediction. We explore two possible sources of sunquakes in the context of the electron beam hypothesis, as an instantaneous transfer of momentum and as a gradual applied force. Recog., IEEE Press, New York, 248. This model comes as a significant improvement to its predecessor in terms of metrics, which are presented in Table4. Y D. Besliu-Ionescu wrote Section 2.1, D.A. In our work, a value of 0.9999 is used for \(\beta \) to halve the effective number of negative samples. The other features are too short-lived to be classified as sunquakes. Given a reference model of the Sun, the differences between its mode frequencies and those of the Sun, if small, are weighted averages of the differences between the Sun's structure and that of the reference model. As a result, the second phase of the experiments is focused on a more recent computer-vision methodology, namely the CL. To better illustrate, if we assume that the Nth frame of an event is present in the training set, then so are all other frames belonging to that same event, and none are present in the validation set. L_{\nu } As with flip operations, rotations are used as a means to upsample our dataset with the purpose of reducing class imbalance. 5% of the images have an erasing rectangle covering 4% of the image. For the complete list and parameters of sunquakes that were positively identified and marked + by Sharykin and Kosovichev (2020) see their Table1. Cutitul de Argint, Nr. Agreement NNX16AC86A, Is ADS down? Findings of this experiment show that the model is only capable of convergence after 100 epochs (top-1 classification accuracy >90.0). The first prediction proved to be a false positive. The \(x\)-axis is represented by the frame index for this event, and the \(y\)-axis denotes the embedding mean value. (CVPR), IEEE Press, New York, 2261. The \(\mathit{TP}\) and \(\mathit{FN}\) labels denote true and false positives, and \(\mathit{TN}\), \(\mathit{FN}\) the true and false negatives. Each mode can then be represented approximately as the product of a function of radius Secondly, when upsampling with geometric transformations, even though they are also applied at runtime to all samples, a transformation bias is induced to the model, making it more inclined to predict sunquakes for geometrically transformed samples. The positions were automatically ingested into the holography method along with a set of other fixed parameters, enabling us to process the sunquake lists and produce temporal cubes of acoustic-egression power corresponding to each event. . Helioseismology, a term coined by Douglas Gough, is the study of the structure and dynamics of the Sun through its oscillations. Cao, Z., Li, X., Feng, Y., Chen, S., Xia, C., Zhao, L.: 2021, ContrastNet: unsupervised feature learning by autoencoder and prototypical contrastive learning for hyperspectral imagery classification. k By subscribing to this BDG newsletter, you agree to our. The datasets generated and/or analyzed during the current study are available in our kaggle sunquakeNet repository: DOI. . Conf. Wiki User 2012-12-10 22:23:11 This answer is: Study guides Incident Command System (ICS) 18 cards When would. Hence, after shuffling, the training and validation set may contain frames belonging to the same initial cube, such as the Nth frame belonging to event 06 April 2001 19:13 residing in the training set, and the \(N+K\)th frame residing in the test set. h This selection is due to the limitations of MDI. Compared to \(L_{2}\), according to Chen, Chen, and Zhang (2019), the logcosh loss improves the reconstruction without damaging the latent-space optimization, thus automatically keeping a balance between the reconstruction and the generation. Our curated positive events for our ML-model training represent only the clearly identifiable and distinguishable sunquakes that were inspected and annotated by us based on the sunquake-detection criteria discussed in Section2.1. \rho Thus, we introduce a dataset constructed from acoustic egression-power maps of solar active regions obtained for Solar Cycles 23 and 24 using the holography method. You can cancel anytime! As the table shows, 7 events are recovered by this model. Before they were able to actually observe a sunquake, astronomers predicted this solar phenomenon could occur. What was the major drawback of this idea? For 30 December 2011 03:03, a significant number of frames are predicted as sunquakes, which, given the high precision of this model, indicates that the dataset might be too different from those previously seen. Per our identification and selection criteria, one and respectively three frames are insufficient to justify a sunquake signature. NASA: Scientists may have unraveled the mystery behind sunquakes - Inverse . The paper presents results of a search for helioseismic events (sunquakes) produced by M-X class solar flares during Solar Cycle 24. Visually, the AR does not seem that clear with respect to the noisy quiet-Sun area around it. However, when using this transform with a 50% probability, the contrastive loss no longer stagnates during training, as it does with other transforms typically used in CL. Because the Sun is a star, helioseismology is closely related to the study of oscillations in other stars, known as asteroseismology. For example, due to current holography limitations, the acoustic sources usually need to be inside the ARs to be identifiable. R It was not until Tom Duvall and Jack Harvey[14] connected the two extreme data sets by measuring modes of intermediate degree to establish the quantum numbers associated with the earlier observations that the higher- Our analysis has shown that many strong sunquakes were produced by solar flares of low M class (M1-M5), while in some powerful X-class flares helioseismic waves were not observed or were weak. The ML methodology is described in detail in Section3, in preparation for Section4, where particularities and limitations of two detection models are presented along with an analysis of current results on both known and tentative newly detected sunquake events. These ripples are essentially acoustic energy and helioseismic holography, in turn, measures the movement of this energy on the Sun's surface. Y Proc. A) Most jovian moons were formed out of their planet's nebula with the same rotational period as their parent planet. What did the gi bill help ex soldiers to do? Secrets Behind Sunquakes Could Lurk Beneath the Solar Surface B) Carbonate materials would form in the oceans more slowly, the atmospheric CO2 content would increase, and the greenhouse effect would strengthen slowly over time. Typically, sunquakes are associated with intense reconnection events in the solar atmosphere resulting in strong solar flares, of X- or M- spectral class, although Sharykin, Kosovichev, and Zimovets (2015) found signatures related to a weaker C.7 class event. Answer: The correct answer is D) Mechanical Wave. These limitations can originate from any combination of instrumental, ML model, interpretation, or statistics effects. They significantly improve our CL-based models performance on our difficult dataset: customized random erase, solarized low pass filter, time based mix. In addition, ground-based networks (notably BiSON and GONG) continue to operate, providing nearly continuous data from the ground too. Figure8 shows clustering of the embeddings produced by the model for the test set. We hypothesize that because of this, the embedding is capable of capturing a gradient in the intensity of the sunquake region between the channels of an individual sample. Understanding the magnitude and evolution of these quakes can help scientists better understand the very nature of solar flares, as well as the conditions of the Sun's interior. Since then, sunquakes have become a more commonly detected phenomenon (Donea, 2011; Kosovichev, 2011; Besliu-Ionescu, Donea, and Cally, 2017; Sharykin and Kosovichev, 2020). We experiment with dimension size for the encoding (latent representation resulting from the backbone) and the projection (latent representation resulting from the contrastive head) varying from 20 and up to 2048 features. 1 / Learn. 2009 IEEE Conf. (or is it just me), Smithsonian Privacy astronomy chapter 14 - Subjecto.com Study ASTRO CHAPTERS 14/15 flashcards. These labels can then be adopted in the CL methods to encourage producing sunquake representations that are dissimilar to their unfounded lookalikes. We attempt to justify this behavior by pinpointing that for this experiment, when training the CL ML model, each input data sample is augmented with the custom time-based mixing transform. Although the solarized low pass filter improves detection metrics, as has been demonstrated in Section3.2 and exemplified in Figure3, we find it is prone to introducing artifact regions into outputs when used in a stand-alone manner for explainability, as it is agnostic of our imposed criteria. Explanation: The guy above me didn't even know Shockwave doesn't belong there. 5 Free The solar neutrino problem referred to the fact that, a couple decades a. In this work, we presented a pedagogical approach on the application of ML methodologies to acoustic emission image data of solar ARs. We model the . This section describes all the data-augmentation methods briefly mentioned in Section3.2, following the categories proposed by (Yang etal., 2022). All these characteristics indicate a typical behavior for immediate pre- and post-quake frames as compared to other event frames, and a more evident discrepancy for sun quake marginal frames. [17] It has long been recognized that measurement of even just a few g modes could substantially increase our knowledge of the deep interior of the Sun. Zhong, Z., Zheng, L., Kang, G., Li, S., Yang, Y.: 2020, Random erasing data augmentation. 10 This enables balanced latent-state representations of the input data, and provides artificially generated reconstructions that follow the learned distributions. While noting that these models are still in a prototype stage, and there is much room for improvement in metrics and bias levels, we hypothesize that their agreement on example use cases has the potential to enable detection of weak solar acoustic manifestations. Although slightly different in distribution, the UMAP is quite consistent in interpretation to the test set clustering shown in Figures5 and 8 discussed above. DOI. AST CH 14 Flashcards | Quizlet The reconstructions for data samples are manually validated to confirm sunquake information is preserved. (eds.) We do this by occluding random areas of both the noisy regions and the AR (see Figure20). To derive a detection model capable of capturing sunquake signatures, several experiments are performed using ML techniques of increasing complexity. Dataset construction: We created a comprehensive dataset of acoustic-emission power maps for ARs that contain sunquake signatures. Based on its surface temperature of 5,800 K, what color are most of the photons that leave the Sun's surface? The data cadence of MDI (SC23) is 60 seconds, while the data cadence of HMI (SC24) is 45 seconds. Furthermore, in AppendicesAD we expand on additional methods and more standard methodological approaches that have been explored while pursuing the task at hand that have not been included in the final proposed solution. [38] These results were subsequently supplemented by analyses that linearize the full set of equations describing the stellar oscillations about a theoretical reference model [18][39][40] and are now a standard way to invert frequency data. (2012) and Sharykin and Kosovichev (2020). Self-supervised (left) and supervised contrastive learning (right) Visualization for sunquake image datasets. Comput. We constructed a curated dataset of major sunquakes during SC23 and SC24, using the holography method. What was the major drawback of this idea? This approach, when applied to contrastive representations, yielded weak results (Macro Avg. DOI. At a first glance, an evident issue with the embedding clustering lies in the distribution of FN samples, which are randomly distributed alongside the TN. Julys Full Buck Moon Will Rise This Week, Gravitational Waves Could Unlock the Answer to the Universes Most Elusive Mystery. B. Zirker in Sunquakes. Neutrinos from the solar core reach the solar surface easily . , Process of transferring data to a storage medium? https://doi.org/10.1007/s11207-022-02081-7, DOI: https://doi.org/10.1007/s11207-022-02081-7. DOI. Autoencoder approach limitations: No autoencoder-based approach, regardless of complexity, proved usable for our dataset. The authors thank Benoit Tremblay for the initial pre-submission review of this work. Kosovichev and Zharkova (1998) first discovered sunquakes as expanding rings in the Dopplergram data using a factor of four image-enhancement technique. The most current models can be provided by us upon reasonable request. In typical CL tasks, occlusion is achieved via the random-crop transform. Helioseismology has shown that the Sun has a rotation profile with several features:[47]. Lindsey, C., Braun, D.C.: 2000, Basic principles of solar acoustic holography (invited review). 1 Kosovichev, A.G.: 2011, Helioseismic response to the X2.2 solar flare of 2011 February 15. , penetrating a characteristic distance Browse the world's largest eBookstore and start reading today on the web, tablet, phone, or ereader. (2022) in order to tackle the imbalance effect. Among the total number of 500 M-X class flares, 94 helioseismic events were detected. He, K., Zhang, X., Ren, S., Sun, J.: 2016, Deep residual learning for image recognition. Astrophys. Article For these four unlabeled events, the predictions are as follows: For 08 May 2012 13:02, one sunquake is predicted, around frames 180188. (eds.) For this reason, before attempting to apply CL and OD techniques described in this article, we decided to use representation learning in the form of unsupervised, or self-supervised, autoencoder approaches. Which is closest to the temperature of the core of the Sun? For instance, 50% of the images have an erasing rectangle covering 1% of the image. Thus, an additional faster region-based CNN (R-CNN) (Ren etal., 2017) based object detection (OD) model is introduced to facilitate interpretation of the model outputs. Consequently it is difficult to infer from them the structure of the solar core. Studies of sunquakes, or helioseismology, have revealed that . We explore the emission of the solar atmosphere during the times indicated by the OD kernel detection. / Sunquakes are seismic emissions visible on the solar surface, associated with some solar flares. Conf. 0/1 points | Previous Answers SeedsFoundations14 8.GR001 My Notes Ask Your Teac Based on what you learned about the solar atmosphere, select all of the correct statements from the following list. DOI. the Sun generates energy by nuclear fusion. In Sunquakes, Jack B. Zirker tells the story of this new science and explains the physics behind these illuminating vibrations. In a subsequent project, including these less certain events will be precisely the focus for fine-tuning the models, after identifying a suitable threshold for applying holography. To further tackle imbalance, besides encapsulating weights in the contrastive loss according to Equation 2, we use SMOTE augmentation (Lematre, Nogueira, and Aridas, 2017) on the training data. The decision to combine the datasets is based on preliminary findings, which indicated that, given the low-data regime, the amount of AR morphologies that an ML model is exposed to needs to be increased such that the model is able to shift the focus from learning representations of ARs to learning sunquake signatures. The framework proposed by Chen etal. What is helioseismology? - Answers [45] For example, if the Sun were rotating uniformly throughout, all the p modes would be split by approximately the same amount. We aimed to enhance the level of explainability of the contrastive model, in the absence of other implemented methods, by additionally utilizing our most impactful augmentation, the solarized low pass filter custom transform, alongside the OD approach. p 10 million K The core of the Sun is hotter and denser than the surface. Cosine distances computed between consecutive frames embedding for the 06 July 2012 13:26 event in the SC23 and SC24 test data, colored by prediction correctness. 2 on Comp. l 39, 1137. Helioseismology, a term coined by Douglas Gough, is the study of the structure and dynamics of the Sun through its oscillations. Found. A compilation of six eight-bit scaled intensity data samples from the ML-prepared dataset. The square of the adiabatic sound speed, Recog. {\displaystyle k_{\rm {h}}\rightarrow \infty } These criteria are further extended in Section2.1. Suppose that large jovian planets had never formed in our solar system. Each of the additional sets should have observation of one sunspot, as retrieved from our literature source (Sharykin and Kosovichev, 2020, Table1). Chen and Zhao, 2021), where a sunquake selection rule can be the conjuncture between impulsive flaring events and downward background oscillatory velocity, occurring at the same location. However, this recent study proved otherwise. Custom augmentations as means of explainability: Although the custom solarized low pass filter augmentation significantly improves both CL models, we find it is not suitable for use a stand-alone tool for ML-output explainability. Local helioseismology is thus somewhat closer in spirit to geoseismology in the sense that it studies the complete wavefield. studies of sunquakes, or helioseismology, have revealed that A) It predicted that the Sun could last only about 25 million years, which is far less than the age of Earth. the solar core has recently revealed its secrets through studies of changing ripples on the Sun's surface and through study of a tiny fraction of mysterious particles generated as part of making the Sun shine . It is similar to geoseismology, or asteroseismology (also coined by Gough), which are respectively the studies of the Earth or stars through their oscillations. (2020) and inspired by the mechanisms of dropout regularization (Yang etal., 2022), is typically used for scenarios where the detection mechanism tends to fail due to occlusions. This information marks a critical step toward a better understanding of the physical mechanics that govern Earth's host star. a final shear layer just beneath the surface, in which the rotation rate slows down towards the surface. 3 An extensive list of representation-learning-centered ML experiments are performed on this dataset, and two end-to-end models are analyzed from a solar-physics phenomenological standpoint. UMAP components constructed from autoencoder-produced embeddings, for all samples in a singular event using 250 (left) and 15 (right) neighbors. The following summary reiterates our most relevant findings and plans for future improvements. In this study, we have applied helioseismic holography, a diagnostic based upon standard wave optics, to reconstruct a 3D image of the sources of acoustic waves emanating from the M9.3-class flare . Further increasing the number of used channels might improve this result. Rods are color blind b. 2023 BDG Media, Inc. All rights reserved. DOI. Due to varying morphological structure between events, reconstructions obtained from latent sizes lower than 1024 do not recover the originals, while larger latent sizes fail in boosting the classification performance, even though they support good reconstructions. Figure19 shows UMAPs built from the latent encoding of sunquake samples. Solar Phys. 6 Suppose that, for some unknown reason, the core of the Sun suddenly became .