Search company, investor...

Stage

Loan | Alive

Total Raised

$1.28M

Last Raised

$680K | 4 yrs ago

About Supercon

Supercon is a company that received a SBIR Phase II grant for a project entitled: A New Production Method for Ta Fibers for Use in Electrolytic Capacitors with Improved Performance and Packaging Options. Their project is intended to develop a new process for manufacturing tantalum (Ta) metal fibers for use in producing tantalum capacitors, and advance this process to the stage of commercialization. This technology, which has been demonstrated in Phase I, could lead to capacitor products having higher performance and greater volumetric efficiency than any currently available. The use of fibers in place of metal powder allows the production of thin anode bodies leading to improved packing options and component performance. The innovation underlying the technology is bundle drawing of Ta filaments in a copper matrix. A composite consisting of Ta filaments in a copper matrix is drawn is a series of reduction steps until the filaments are less than about 10 microns in diameter. The drawn wire is rolled to produce ribbon-type filaments that are 1 micron or less in thickness. The copper composite matrix is chemically dissolved without attacking the Ta to produce metallic Ta high surface area, ribbon-fibers. The fibers are formed into thin mats, which are sintered to produce porous metal strips from which high surface area capacitor anodes are made. A significant aspect of this approach is that fiber morphology can be varied over a wide of fiber thicknesses unlike powder. This allows the morphology of the fibers to be optimized for the particular voltage rating and use requirements in order to maximize the performance of the capacitor. Commercially, nearly all medical, automotive, military and many consumer electronic devices utilize Ta electrolytic capacitors due to their outstanding performance, reliability and volumetric efficiency. Solid electrolytic capacitors are currently made from Ta metal powder. Several million pounds per year of Ta powder are consumed in manufacturing Ta capacitors for these applications. The trend in electronics is toward high powder components and increased miniaturization. Combined with the need to lower materials and manufacturing costs, these considerations have created an opportunity for new method of producing solid electrolytic capacitors. Fiber metal technology has the potential to both lower manufacturing costs, improve capacitor performance, and improve packaging options, which could enable the development of new product that are either currently very difficult or very expensive to make using current technology base on metal powder. Supercon is a company that received a SBIR Phase I grant for a project entitled: A New Production Method for Ta Fibers for Use in Electrolytic Capacitors with Improved Performance and Packaging Options. Their project is intended to demonstrate a new process for manufacturing valve metal fibers for use in producing capacitors. The technology is applicable to all valve metals used for making solid electrolytic capacitors. If successful, this technology could lead to capacitor products having higher performance and greater volumetric efficiency than are currently available. The use of fibers in place of the standard powder compacts allows the production of thin anode bodies leading to improved packaging options and component performance. The innovation underlying the technology is bundle drawing of valve metal filaments contained in copper matrix. A composite consisting of valve metal filaments in a copper matrix is drawn in series of reduction steps until the filaments are less than 10 microns. The drawn wire is rolled to produce submicron thick ribbon type filaments. The copper composite matrix is chemically dissolved to produce metallic thin fibers. The fibers are formed into thin mats, which are sintered to produce porous metal strips from which high surface area capacitor anodes can be made. A significant aspect of this approach is that fiber morphology can be varied within a wide range of thickness and widths unlike powders. This allows the morphology of the fibers to be optimized in order to maximize the properties of the capacitor. Commercially, nearly all medical, automotive and consumer electronic devices all utilize solid electrolytic capacitors due to their performance, volumetric efficiency, and high reliability. Several million pounds per year of powder are consumed in the manufacture of capacitors for these applications. The trend towards higher power components, and miniaturization, combined with the need to lower materials and manufacturing costs have created an opportunity for new methods of producing solid electrolytic capacitors. Fiber metal technology has the potential to both lower manufacturing costs, improve capacitor performance, and improve packaging options which could lead to new products that are either very difficult or very expensive to make using current methods.

Headquarters Location

830 Boston Tpke

Shrewsbury, Massachusetts, 01545,

United States

(508) 842-0174

Loading...

Loading...

Supercon Patents

Supercon has filed 1 patent.

patents chart

Application Date

Grant Date

Title

Related Topics

Status

12/18/2008

1/25/2011

SAGE sites, Food safety, Reliability engineering, IBM mainframe operating systems, Process management

Grant

Application Date

12/18/2008

Grant Date

1/25/2011

Title

Related Topics

SAGE sites, Food safety, Reliability engineering, IBM mainframe operating systems, Process management

Status

Grant

Latest Supercon News

Predicting superconducting transition temperature through advanced machine learning and innovative feature engineering

Feb 17, 2024

Abstract Superconductivity is a remarkable phenomenon in condensed matter physics, which comprises a fascinating array of properties expected to revolutionize energy-related technologies and pertinent fundamental research. However, the field faces the challenge of achieving superconductivity at room temperature. In recent years, Artificial Intelligence (AI) approaches have emerged as a promising tool for predicting such properties as transition temperature (Tc) to enable the rapid screening of large databases to discover new superconducting materials. This study employs the SuperCon dataset as the largest superconducting materials dataset. Then, we perform various data pre-processing steps to derive the clean DataG dataset, containing 13,022 compounds. In another stage of the study, we apply the novel CatBoost algorithm to predict the transition temperatures of novel superconducting materials. In addition, we developed a package called Jabir, which generates 322 atomic descriptors. We also designed an innovative hybrid method called the Soraya package to select the most critical features from the feature space. These yield R2 and RMSE values (0.952 and 6.45 K, respectively) superior to those previously reported in the literature. Finally, as a novel contribution to the field, a web application was designed for predicting and determining the Tc values of superconducting materials. Introduction The amazing properties of superconducting materials are a direct consequence of quantum mechanics that emerge on a large scale 1 . The two basic characteristics of superconductors that make this class of materials different from others include: a) offering no resistance to the flow of electric currents, and b) complete exclusion of magnetic field 2 . No comprehensive theory capable of predicting the transition temperatures (Tc) of superconducting materials has yet been presented to date and the discovery of new superconductors still relies on expert intuition and is largely dependent on trial and error based on experience 3 . Hence, empirical laws have for many years served as guides for researchers in their efforts to fabricate new superconducting materials 4 . Condensed Matter Physics strives to discover the interactions of materials at the atomic level since material properties are derived from these interactions 5 . Prediction and determination of the microscopic properties of materials presuppose the solution of the Schrodinger equation for a Many-Body system. However, solving this equation for such systems is practically impossible due to the vast Hilbert space needed to handle them, especially for highly correlated materials. Consequently, a solution adopted in most cases is to employ approximate methods 6 , 7 , 8 . One of these methods is Density Functional Theory (DFT) which is based on the Hohenberg–Kohn and Kohn–Sham theorems and has a substantial record of success in predicting material properties and solving the associated quantum mechanics problems 9 , 10 , 11 . Despite its outstanding achievements, the theory has some limitations in its current form; for instance, it employs approximation for exchange–correlation functional, yields errors when used for strong correlation systems, can only be employed for a small number of atoms, and is hampered by increasing computational costs and runtime with increasing system size 10 , 12 , 13 , 14 , 15 . Strong electron–electron correlations in superconducting materials make it extremely challenging to perform first-principles calculations to determine their structural properties and predict their Tc 3 , 9 , making searching for novel alternative approaches inevitable. As alternative strategies for solving quantum mechanics problems, machine learning methods offer lower computation costs, shorter execution times, accurate predictions, and faster development cycles 9 , 12 , 13 . Being data-driven and given the fact that huge amounts of data have been produced over the years, machine learning methods encourage researchers to utilize them for discovering novel materials and predicting their properties 4 , 5 , 6 , 16 . Materials Science is nowadays said to have entered its fourth stage of evolution, termed “Data-Based Materials Science”, a term borrowed from Thomas Samuel Kuhn to describe the field’s development 6 , 12 , 17 , 18 . Figure  1 illustrates the four (empirical, theoretical, computational, and data-driven) paradigms of materials science. To date, large amounts of theoretical and experimental data have been collected in the three traditional (i.e., empirical, theoretical, and computational) paradigms; the next step, logically, is to apply the new innovative tools developed by artificial intelligence, which are capable of extracting knowledge from such data 6 , 12 , 18 , 19 , 20 , 21 , 22 . Figure 1 Given the importance of the Tc values of superconducting materials, researchers have in recent years developed machine learning-based models for predicting this quantity. Selecting 21,263 superconducting materials and utilizing 80 atomic descriptors for each compound, Hamidieh 4 used the XGBoost algorithm to design a model for predicting of Tc. Stanev et al. 16 employed the Random Forest algorithm to develop a model using 132 atomic features of Magpie descriptors for 6196 superconducting compounds. Konno et al. 3 implemented a convolutional neural network (CNN) model (i.e., a deep learning model) to predict the Tc values of about 13,000 superconducting materials. They represented their materials using an innovative “periodic table reading” method. The dimensions of the representation were 4 × 32 × 7, with 4 representing the four orbitals of s, p, d, and f corresponding to the valence electrons of each element in a compound, and 32 and 7 denoting the dimensions of the periodic table. Dan et al. 23 developed the ConvGBDT model by merging the convolutional neural network (CNN) and the gradient boosting decision tree (GBDT) models. For the three datasets of DataS, DataH, and DataK, the authors used the Magpie descriptors to represent materials and the ConvGBDT model to predict Tc values. Li et al. 11 introduced a hybrid neural network (HNN) model as a combination of a convolutional neural network (CNN) and a long short-term memory neural network (LSTM). They utilized atomic vectors and employed both the one-hot and Magpie material characterization methods to represent superconductors in the feature space. The authors found that the Magpie features generally outperformed the one-hot features. Roter et al. 24 employed the Bagged Tree method (a variant of the Random Forest algorithm) to design a model for predicting Tc. They represented superconducting materials using a chemical composition matrix as the feature space. The matrix had about 30,000 rows and 96 columns, wherein each row corresponded to a chemical formula, and the columns contained the 96 primary elements of the periodic table. Each entry in this matrix was filled with an index corresponding to the elements of each chemical compound. Quinn et al. 25 utilized a Crystal Graph Convolutional Neural Network (CGCNN) model to integrate classification and regression models within a pipeline to identify candidates of high-temperature superconductors from among the 130,000 compounds in the Materials Project. In the crystal-graph representation of materials, the connections between atoms represent the graph’s edges, and the locations of the atoms and their properties represent the vertices. The main objective of the current research is to design a suitable and reliable model for predicting the Tc values of superconducting materials using machine learning approaches. While the algorithm and the dataset are the two indispensable research tools in data science, the present study attaches more importance to the dataset than the algorithm. After carefully cleaning data, we generate a suitable feature space for superconducting materials. The main advantages of the present work over previous ones include: (1) Establishing more appropriate feature space related to superconducting Tc and (2) Identifying the features most related to the Tc values of superconducting materials. We reach significant results by designing the Jabir package to produce 322 atomic features for each compound and Soraya package for selecting features. Data Data set Two essential steps must be taken before statistical learning can predict Tc in superconducting materials. The first involves collecting and preprocessing a dataset, and the second is adopting a suitable algorithm for the learning process and model development on that dataset. According to Halevy et al. 26 , the first step is of greater significance as data scientists typically devote about 80% of their efforts to datasets and their preprocessing 27 ; the same is valid with the present work using SuperCon dataset ( https://doi.org/10.48505/nims.3739 ), currently the largest and most comprehensive superconducting materials database containing 33,407 superconducting compounds. Here, a significant contribution is done by executing distinct steps of data pre-processing and providing detailed explanations for each step. Ultimately, following the implementation of various data pre-processing phases and the exclusion of problematic data, the DataG dataset consisting of 13,022 superconducting compounds is derived. Cleaning the dataset Dealing with missing and duplicated data The SuperCon dataset lacks the transition temperature values for 7088 compounds. These cases are identified as missing data and removed from the dataset. Along with that, we remove 7418 data duplications. Among these, 1264 compounds are regarded as duplicates due to the displacement of data elements; examples include: MgB2, B2Mg, Ag7B1F4O8, Ag7F4O8B1, Al0.1Si0.9V3, V3Si0.9Al0.1, Zr2Co1, Co1Z2, …. Dealing with problematic data (1) We eliminate 5348 compounds whose element subscripts are X, Y, Z, D, x, y, z, and d. (2) We remove problematic compounds such as: HgSr2Ho0.333Ce0.667Cu2O6=z, Ba2Cu1.2Co2.4O2,4, Ag7Bf4O8, Hg0.3Pb0.7Sr1.75La0.25CuO4+2, Ho0.8Ca0.2Sr2Cu2.8P0.2Oz+0.8, Bi1.6Pb0.4Sr2Ca2Cu3F0.8Oz-0.8. (3) Compounds containing the elements not included in the periodic table are ignored. (4) Given the objective of predicting transition temperatures for superconductors at ambient pressure, those created under non-ambient pressures (e.g., La1H10, H2S1, H3S1, D3S1, …) are removed from the dataset. (5) The compound YBa2CuO6050 is eliminated on the grounds that the oxygen subscript of 6050 might be incorrect 4 . (6) We dismissed 70 compounds whose transition temperatures are reported to be zero. (7) Finally, the compound Pb2CAg2O6 is discarded due to the unreasonable transition temperature of 323 K reported for this compound. Data correction (1) According to the SuperCon reference 28 , the transition temperature of the iron-based superconductor CsEuFe4As4 is nearly 30 K, while the SuperCon dataset records it as 287 K. Therefore, it is modified to 28.7 K. Moreover, the compound Sm1Ba-1Cu3O6.94 is substituted with Sm1Ba1Cu3O6.94. (2) Bi1.6Pb0.4Sr2Cu3Ca2O1013 is altered to Bi1.6Pb0.4Sr2Cu3Ca2O10.13 because the nearby data rows containing formulas with O10.xx 4 . Dealing with multiple temperatures reported for a single compound One limitation in the SuperCon dataset is the presence of multiple Tc values reported for 2132 compounds, posing a challenge for accurate analysis. For instance, MgB2 alone has been reported to exhibit 47 different transition temperatures ranging from 5 to 40.5 K. To tackle this challenge, it has been recommended to consider average transition temperatures for compounds that have multiple Tc values reported in the dataset. Prior to determining the average Tc value, it is essential to exclude compounds whose reported transition temperatures display significant dispersion. To achieve this, the standard deviation of the different transition temperatures for each compound is calculated and compounds with standard deviations greater than 20 K are removed from the dataset. Performing this procedure leads to the elimination of 18 compounds. Detecting outliers Undoubtedly, outliers in a dataset can pose problems in identifying underlying patterns, resulting in diminishing system performance and accuracy 29 . In this study, the outlier data are detected using the Z-score method 30 and the PyOD package 31 , both renowned tools in the field of anomaly detection. After a meticulous examination, the outliers are identified and excluded according to the three following distinct aspects: (1) Transition temperature: The average transition temperature of remaining compounds ranges from 0.0005 to 250 K. Using the abovementioned techniques, 10 superconducting materials with average transition temperatures outside the 0.01–136 K range are identified as outliers and removed. Figure  2 illustrates the Tc distribution of the few superconducting material families. (2) Number of elements: Fig. 3 shows the number of compounds according to the number of constituent elements. A subset of compounds with one, eight, and ten elements are identified as outliers and subsequently removed, resulting in the elimination of 81 superconducting compounds. (3) The summation of subscripts: Implementing the abovementioned techniques reveals that six compounds exhibited subscript summations exceeding 100 that are subsequently removed as outliers. Figure 2 Computational methods Machine learning algorithm In this study, we use the CatBoost algorithm as a machine learning ensemble technique based on Gradient Boosted Decision Trees (GBDT) proposed by Yandex Company. GBDT is an efficient tool for solving regression and classification issues in big data sets. CatBoost is a Decision Tree based algorithm and open-source implementation for supervised machine learning that involves two innovations: Ordered Target Statistics and Ordered Boosting. Researchers have successfully employed CatBoost for machine learning investigations incorporating Big Data since its launch in late 2018. Numerous applications have been reported for CatBoost in various fields, including astronomy, finance, medicine, biology, electrical utilities fraud, meteorology, psychology, traffic engineering, cyber-security, biochemistry, and marketing 32 . However, the application of CatBoost has not yet been reported for predicting superconducting transition temperatures. This study uses the algorithm to find if it can efficiently identify relationships and patterns between features and Tc. We show that through the creation of atomic features for superconducting material, CatBoost algorithm provides a model with very good accuracy. Generating the feature space After preprocessing the data set, we must extract atomic features in a “data representation” procedure. There are two main approaches for representing compounds: The first is based on chemical formulas, and the second on crystal structure 23 . The atomic features are generated for superconducting materials using the first approach for our purposes. In fact, machine learning algorithms recognize a compound by its characteristics, i.e. the identifier and characteristic of a compound are the features that are consider for the compound. This process is called data representation. Figure  4 shows how to calculate atomic features. Figure 4 Conclusion In the realm of materials science, artificial intelligence stands as a powerful tool for predicting material properties. In this study, the CatBoost algorithm was employed to predict the Tc values of superconducting materials, marking a novel approach. For this purpose, data pre-processing of the SuperCon dataset was accomplished as a significant step in data science to develop a new dataset called DataG containing 13,022 superconducting compounds. Also, a new Jabir package capable of generating 322 atomic descriptors was designed and developed. Comparisons revealed the superiority of the atomic features generated by Jabir over those generated by such previous ones as the Magpie package. Furthermore, an innovative hybrid technique was developed as the feature selection method (Soraya package). In order to design and develop Jabir and Soraya packages, we applied novel ideas and innovative approaches, such as: (i) using new and diverse physical atomic features in the Jabir package and considering three different states (Elemental, Subscript, Fraction) in order to calculate the atomic features of each compound and (ii) using an innovative hybrid technique in Soraya package, removing features that are highly correlated with each other (removing redundant features) and using SHAP's technique to select the most important features and finally using the forward method to adding the most important features. The contributions of the study led to optimized evaluation values (R2, RMSE, MAE) of DataH, DataS, and DataK datasets without the need for any data pre-processing. The present study’s results indicate that the procedure of selecting the most important descriptors significantly impacts predicting superconducting materials’ Tc values. Finally, the development of a novel web application was a pioneering contribution to the field for predicting and determining the Tc of superconducting materials. Data availability The dataset (DataG), which is prepared after various steps of data pre-processing on the SuperCon dataset, is available at the following address. https://github.com/Gashmard/DataG_13022_superconducting_materials Code availability The developed packages (Jabir and Soraya) and the web application are accessible at the following URLs. Web application: https://supercon-tc.iut.ac.ir/ Supplementary Information 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/ .

Supercon Frequently Asked Questions (FAQ)

  • Where is Supercon's headquarters?

    Supercon's headquarters is located at 830 Boston Tpke, Shrewsbury.

  • What is Supercon's latest funding round?

    Supercon's latest funding round is Loan.

  • How much did Supercon raise?

    Supercon raised a total of $1.28M.

  • Who are the investors of Supercon?

    Investors of Supercon include Paycheck Protection Program, U.S. Department of Energy and National Science Foundation.

Loading...

Loading...

CBI websites generally use certain cookies to enable better interactions with our sites and services. Use of these cookies, which may be stored on your device, permits us to improve and customize your experience. You can read more about your cookie choices at our privacy policy here. By continuing to use this site you are consenting to these choices.