Invented by Jiho Yoo, Seokho KANG, Youngchun KWON, Kyung doc KIM, Jaikwang SHIN, Hyosug LEE, Younsuk Choi, Samsung Electronics Co Ltd

The market for methods and apparatus for discovering new materials using machine learning on targeted properties is rapidly growing and holds immense potential for various industries. With advancements in technology and the increasing demand for innovative materials, machine learning has emerged as a powerful tool to accelerate the discovery and development of new materials with desired properties. Traditionally, the process of discovering new materials has been time-consuming, expensive, and often reliant on trial and error. Researchers would conduct experiments and analyze the results to identify materials with specific properties. However, this approach is limited by the vast number of possible combinations and the complexity of materials’ behavior. Machine learning, on the other hand, offers a data-driven approach to material discovery. By training algorithms on large datasets of existing materials and their properties, machine learning models can learn patterns and relationships that enable them to predict the properties of new materials. This allows researchers to narrow down the search space and focus on materials with the highest likelihood of exhibiting desired properties. The market for methods and apparatus for discovering new materials using machine learning is driven by several factors. Firstly, the demand for advanced materials with specific properties is increasing across industries such as energy, electronics, healthcare, and aerospace. For example, the renewable energy sector requires materials with improved efficiency and stability for solar panels and energy storage devices. Machine learning can help identify materials that meet these requirements, leading to significant advancements in clean energy technologies. Secondly, the cost and time savings offered by machine learning-based material discovery are highly attractive. Traditional experimental approaches can take years and require substantial resources. Machine learning algorithms can accelerate the process by suggesting promising materials for further investigation, reducing the time and cost associated with trial and error experiments. Moreover, machine learning can enable the discovery of materials with properties that may not have been previously considered. By analyzing vast amounts of data, algorithms can identify complex patterns and correlations that human researchers may overlook. This opens up new possibilities for materials with unique and desirable characteristics, leading to innovation in various industries. The market for methods and apparatus for discovering new materials using machine learning is witnessing significant growth and competition. Numerous startups and research institutions are developing and commercializing machine learning platforms specifically designed for material discovery. These platforms combine advanced algorithms, high-performance computing, and extensive databases to provide researchers with powerful tools for accelerating their material discovery efforts. Additionally, collaborations between academia, industry, and government agencies are driving the market forward. These partnerships leverage the expertise of researchers, the resources of industry, and the support of government funding to advance the field of machine learning-based material discovery. This collaborative approach fosters innovation and ensures the development of robust and reliable methods and apparatus for discovering new materials. In conclusion, the market for methods and apparatus for discovering new materials using machine learning on targeted properties is poised for significant growth. The demand for advanced materials with specific properties, coupled with the cost and time savings offered by machine learning, is driving the adoption of this technology across industries. As machine learning algorithms continue to improve and datasets expand, the discovery of new materials with tailored properties will become increasingly efficient and transformative for various sectors.

The Samsung Electronics Co Ltd invention works as follows

The invention is a “structure-generating method” for generating structure candidates of new materials. It includes: a structure-generating process performing machine-learning on a machine-learning model, which is configured to give a result on the basis of a descriptor, a physical characteristic of a material and its structure; and generating structure candidates of new materials based on the results of machine-learning, where the new material has the target physical property and the descriptor, physical property and structure are all stored in

Background for Methods and apparatus for discovering new materials using machine learning on targeted properties

(1) Field of the Invention.

This disclosure is a method for creating a structure-candidate of a novel material.

(2) Description of Related Art

First principles calculations can be used to predict material characteristics without direct measurements or related experiments. As high-performance computing techniques using supercomputers develop, the evaluation and screening of a range of material candidates will be possible using a high throughput computational method. This is quicker and more efficient than direct synthesis and characterisation of the array using conventional methods.

A screening range is determined for high-throughput computation screening method. Researchers can select the screening search range based on criteria for determining materials in a database or by using a rule to modify or combine existing materials. The success or failure of high-throughput computation screening is dependent on the quality of the search range. However, the search range can be determined by a researcher based on their experience and intuition.

Material development techniques aim to create a material with performance that is significantly better than existing materials. It may be necessary to perform multiple computational screenings and search range settings to achieve the desired performance. When a target material is not found within a predetermined search range, a drawing is made based on an analysis of the existing screening results. The new search region is then determined using the drawn idea. In this case, however, it’s difficult to analyze and determine the next search region because of a large number of calculation data.

The invention provides a method of generating a candidate structure for a new material with a desired physical property based on the results of machine learning performed on a descriptor and a physical property of an existing material.

An embodiment” provides an apparatus that generates a structure candidate for a new material with a target physical properties based on the result of machine-learning that is performed on a descriptor of a material, its physical property and structure.

An embodiment” provides a method of generating a candidate structure for a new material. A structure-generating processor may perform machine learning using a machine-learning model. The machine learning model can be configured to produce a result that is based on the descriptor, physical property, and structure of a particular material. The structure-generating process may also include generating, by the structure generating processor, an initial structure candidate for the new material. The new material is characterized by a desired physical property. A database stores the description of the new material, its physical property and its structure.

The structure-generating processor may perform the machinelearning on a relation between the physical property and descriptor, as well as the machinelearning on a relation between the factor to the structure.

The structure-generating processor may determine a structural factor that indicates the relationship between factor and structure when performing machine learning.

The structure-generating processor may determine the structural factors for each of the layers of the plurality.

The structure-generating processor may determine the factor, for example, by learning an encoding functions to derive the factor from the description based on the data related to both the descriptors and the physical properties; and by applying the encoding functions to the descriptor.

The structure-generating processor learns a decoding algorithm for generating a structure candidate from analyzing the data of each factor and structure.

The structure-generating processor may generate the structure candidate, for example, by sampling a factor that corresponds to the physical property target; and applying the decoding functions to the factor.

The structure-generating processor may also learn a prediction function to predict the physical properties from the factor, based on each descriptor and physical property.

The structure-generating processor can also generate an arbitrary description, predict a physical characteristic of a material that corresponds to the descriptor using the prediction function and determine whether the predicted physical characteristic of the material that corresponds to the descriptor is the target physical quality.

The generating of structure candidates may include, by the structure-generating processor: when the predicted property of the materials corresponding the arbitrary description does not have target physical properties, generating an arbitrary descriptive different from the original arbitrary description; predicting the physical property a material by applying the prediction function on the different descriptor, and determining if the predicted property of material corresponding the new arbitrary description has the target property.

The generating of a structure candidate can further include the following: by applying an encoder function to derive the factor that corresponds to the descriptor, determining the factor for the descriptor, and by applying a decoding to generate the structure from the factor.

The structure-generating process may also include performing a valid check and a redundant check on the candidate structure corresponding to the descriptor arbitrarily chosen.

A structure-generating device is provided for generating structure candidates of new materials. The structure-generating device may include a database configured for storing a descriptor and physical property of a new material. It can also include a processor that is configured to perform machine-learning on a model of machine-learning, which is configured using the descriptor and physical property.

Click here to view the patent on Google Patents.