| | Recent milestones in the development of artificial intelligence (AI) have taken place in areas with high or unlimited data availability. Examples include speech processing, image processing, and game AI. While for language and images enormous amounts of labeled training data are available (e.g. each solution of a graphical Google-Capcha contributes a new data set), in games the repeated play of two AIs against each other can generate almost unlimited training data. In contrast, the effort in natural sciences and especially in material sciences is incomparably higher. The collection of a single dataset can take hours to weeks and cause enormous costs. The purely data-driven approaches that have been so successful up to now can therefore not be readily applied in this field. Instead, ways must be found to significantly reduce the physical data requirements (process data, measurement data) of AI methods. One way to do this is to generate virtual data using simulation models, the use of which is much more scalable and easier to automate than data collection in the laboratory. However, the availability of simulation models varies widely across material classes. While powerful ab initio methods are available for metals, for example, to calculate macroscopic properties on the basis of element composition, the modeling of polymerization reactions is still very challenging and characterized by a high degree of fuzziness. Even more difficult is the simulation of complex systems such as a battery, in which various chemical and physical effects interlock over several scales. In these areas in particular, the inclusion of explicit knowledge is therefore additionally required to reduce data requirements. This enables AI to include explicit relationships between data in addition to data (hybrid AI). Explicit relationships can be, for example, unambiguous relationships such as chemical constituents or physical laws, but also weaker forms such as an assumed but unquantifiable relationship between parameters. However, this requires that the explicit knowledge is expressed (formalized) in a machine-readable form. This is exactly what ontologies are able to do by transferring the almost unlimited complexity of human knowledge into a strict description language without loss of information. At the same time they enforce a standardization, which is necessary for the multidisciplinary consolidation of explicit knowledge. The knowledge contained in BattInfo and BVCO about the internal relationships in the battery and the external relationships in its manufacturing can therefore help to significantly reduce the data requirements in future AI applications. | | Recent milestones in the development of artificial intelligence (AI) have taken place in areas with high or unlimited data availability. Examples include speech processing, image processing, and game AI. While for language and images enormous amounts of labeled training data are available (e.g. each solution of a graphical Google-Capcha contributes a new data set), in games the repeated play of two AIs against each other can generate almost unlimited training data. In contrast, the effort in natural sciences and especially in material sciences is incomparably higher. The collection of a single dataset can take hours to weeks and cause enormous costs. The purely data-driven approaches that have been so successful up to now can therefore not be readily applied in this field. Instead, ways must be found to significantly reduce the physical data requirements (process data, measurement data) of AI methods. One way to do this is to generate virtual data using simulation models, the use of which is much more scalable and easier to automate than data collection in the laboratory. However, the availability of simulation models varies widely across material classes. While powerful ab initio methods are available for metals, for example, to calculate macroscopic properties on the basis of element composition, the modeling of polymerization reactions is still very challenging and characterized by a high degree of fuzziness. Even more difficult is the simulation of complex systems such as a battery, in which various chemical and physical effects interlock over several scales. In these areas in particular, the inclusion of explicit knowledge is therefore additionally required to reduce data requirements. This enables AI to include explicit relationships between data in addition to data (hybrid AI). Explicit relationships can be, for example, unambiguous relationships such as chemical constituents or physical laws, but also weaker forms such as an assumed but unquantifiable relationship between parameters. However, this requires that the explicit knowledge is expressed (formalized) in a machine-readable form. This is exactly what ontologies are able to do by transferring the almost unlimited complexity of human knowledge into a strict description language without loss of information. At the same time they enforce a standardization, which is necessary for the multidisciplinary consolidation of explicit knowledge. The knowledge contained in BattInfo and BVCO about the internal relationships in the battery and the external relationships in its manufacturing can therefore help to significantly reduce the data requirements in future AI applications. |
| | In addition to the actual development, extension, and merging of ontologies, making them usable is a future challenge. Researchers worldwide need to be enabled to semantically express both their research data and their findings using ontogies without being ontology experts themselves. For this purpose, graphical user interfaces automatically generated from the ontologies and semi-automated conversion tools are needed. | | In addition to the actual development, extension, and merging of ontologies, making them usable is a future challenge. Researchers worldwide need to be enabled to semantically express both their research data and their findings using ontogies without being ontology experts themselves. For this purpose, graphical user interfaces automatically generated from the ontologies and semi-automated conversion tools are needed. |
