Materials Informatics (MI) is an approach that accelerates the discovery and design of new materials using data science methods, including machine learning.In November 2023, Google DeepMind announced that it had discovered 2.2 million new crystal structures using its deep learning tool “GNoME” (published in *Nature*),and in January 2025, a paper on Microsoft’s generative AI “MatterGen” was published in *Nature*; as such, AI-driven materials discovery has been attracting significant attention in recent years.In Japan as well, a joint venture between Preferred Networks (PFN) and ENEOS has been offering the general-purpose atomic-level simulator “Matlantis” via the cloud since July 2021, and as of August 2024, it has been adopted by more than 90 companies and organizations.
In this article, after verifying the existence of patents in this field using primary sources (Google Patents, J-PlatPat, and WIPO PATENTSCOPE), we examine: (1) statistics on patent application trends; (2) eight registered patents by Japanese companies; (3) six registered patents by foreign companies;④ the core of examination practice regarding whether patents can be granted for “materials predicted solely by AI” (Japan Patent Office AI examination cases 51 and 52, among others), and ⑤ the appropriate use of patents versus trade secrets—all from the perspective of a patent attorney.
Table of Contents
Traditional materials development has centered on hypothesis formulation based on researchers’ experience and intuition, followed by repeated trial and error through experimentation, meaning it has taken many years to bring new materials to practical use.MI aims to significantly reduce this trial-and-error process by using machine learning to extract correlations between material structure and properties from accumulated experimental and computational data, thereby predicting “which candidate should be tested next.”The “Strategy for Strengthening Material Innovation Capabilities” (adopted on April 27, Reiwa 3; revised on June 4, Reiwa 7) by the Council for the Promotion of Integrated Innovation Strategy also establishes the development of a data-driven R&D infrastructure as a fundamental national policy.
The following three events can be cited as technological turning points:
| Date | Event |
|---|---|
| July 6, 2021 | The PFN×ENEOS joint venture (PFCC) began offering “Matlantis,” a general-purpose atomic-level simulator. It claims to be tens of thousands of times faster than first-principles calculations (DFT) thanks to deep learning. In September 2024, the company announced support for 96 elements and adoption by more than 90 companies and organizations. |
| November 29, 2023 | Google DeepMind announced that it had discovered 2.2 million new crystal structures (of which approximately 380,000 are considered the most stable) using “GNoME” (published in *Nature*). According to external research, 736 of these have already been independently synthesized experimentally. |
| January 16, 2025 | Microsoft’s “MatterGen” paper was published in *Nature*. The study demonstrated the direct generation of inorganic materials with desired properties using a diffusion model and validated the experimental synthesis of the generated material TaCr₂O₆ (bulk elastic modulus: measured at 169 GPa compared to the designed value of 200 GPa). |
There are significant differences in market size definitions among research firms; Grand View Research estimates approximately $135 million in 2023 (with a compound annual growth rate of 16.5%),while 360iResearch estimates approximately US$720 million by 2025 (CAGR of 18.34%); although there is a range in these estimates, they all agree that high growth rates are expected.
The Japan Patent Office highlighted MI in its Survey on Technological Trends in Patent Applications for Reiwa 1 (published in February Reiwa 2). According to the survey, there were a total of 493 MI-related patent families in Japan, the U.S., Europe, China, South Korea, and India from 2010 to 2017 (based on the year of priority claim).By applicant nationality, China stood out with 332 applications (67.3%), followed by South Korea with 55 (11.2%), Japan with 37 (7.5%), and the United States with 35 (7.1%).Applications began to increase around 2015 (48 in 2015 → 116 in 2017), a trend that coincided with the deep learning boom. Among Japanese applicants, Pacific Cement (10 applications) ranks among the top.
Note: These statistics were compiled in 2020 based on applications filed between 2010 and 2017; the most recent situation has changed further. For reference, the Japan Patent Office’s “Survey on the Status of AI-Related Invention Applications”(October 2025 edition), applications for AI-related inventions in Japan increased to approximately 11,400 in 2023, while applications for core AI inventions (classified under G06N) in China exceeded 100,000 in 2023—more than five times the number in the United States.
In terms of patent classifications, the January 2019 revision of the International Patent Classification (IPC) introduced the new “G16C (Computational Chemistry, Chemical Informatics, and Computational Materials Science)” class, which includes groups such as machine learning and chemometrics (G16C 20/70),prediction of physical properties of compounds (G16C 20/30), molecular design (G16C 20/50), and computational materials science (G16C 60/00).In practice, MI-related inventions are often searched for and classified using a combination of this G16C (application-side) and G06N (AI methodology-side), which represents machine learning itself; covering these two categories serves as the starting point for prior art searches and FTO analyses.
Below, we present eight registered patents from Japanese companies for which our firm verified the bibliographic data and registration status directly on the Google Patents, J-PlatPat, and WIPO PATENTSCOPE websites (survey conducted as of July 18, 2026; this is not an exhaustive list).
| Patent Number | Patentee | Key Technical Points |
|---|---|---|
| Patent No. 6950119 | Showa Denko (now Resonac) | A “division-of-labor” materials design system in which data scientists create models and a large number of non-experts use them (supports both forward and inverse problems) |
| Patent No. 7190615 | Showa Denko (now Resonac) | A method for predicting temperature-dependent material properties (such as the tensile strength of aluminum alloys) by adding the evaluation temperature and holding time of the material properties as explanatory variables |
| Patent No. 7109339 | Showa Denko | Calculates predicted values and uncertainties of physical properties using Gaussian process regression, then recommends polymer structures to be verified (Bayesian optimization-based search) |
| Patent No. 7218519 | Yokohama Rubber | Predicts physical properties of vulcanized rubber using machine learning based on compounding and processing conditions. Features preprocessing that uses regression analysis to interpolate data for items with few samples |
| Patent No. 7215017 | Yokohama Rubber | Reverse search for rubber compounds that achieve target physical properties using a trained model and evolutionary algorithms (subject to constraints on the number of raw materials) |
| Patent No. 7302297 | Fujitsu | Predicting dielectric constant and other properties by calculating molecular structural similarity using “maximum independent set search” |
| Patent No. 7388217 | Fujitsu | Plots crystal structures as graphs, formulates similarity analysis as an Ising model, and solves it using an annealing machine (quantum-inspired computing × MI) |
| Patent No. 7403032 | Preferred Networks | Generalization technology that incorporates differences in first-principles calculation conditions as “label information” during the training of Neural Network Potentials (NNP) |
To illustrate how MI patent claims are typically drafted, I will cite Claim 1 of Patent No. 6950119 as a representative example.
“A material design system for designing target design materials, including materials composed of multiple compositions or materials manufactured through combinations of multiple manufacturing conditions,an expert terminal capable of utilizing a model training interface for performing machine learning on a model that takes the correspondence between the design conditions of the target material and the material property values as input and output; and for a specific target material,for a specific target material, a material design interface capable of estimating the material property values from the design conditions, or the design conditions from the material property values, using a trained model created by the expert terminal for said specific target material; and
It is noteworthy that this claim does not focus on the machine learning algorithm itself, but rather defines the invention in terms of the system configuration (operational mechanism) consisting of the “model creator (expert terminal)” and the “model user (multiple general-purpose terminals).”Even if it is difficult to differentiate the predictive model itself, incorporating it into a mechanism for running MI in the field of materials development allows for patent protection—this can be considered a typical approach for MI patents.
PFN’s Patent No. 7403032 (registered December 21, 2023), which is classified as part of Matlantis’s core technology, relates to a method for training NNP (Neural Network Potential).Since first-principles calculations yield results that vary slightly depending on the software and computational parameters used, simply mixing training data generated under different conditions results in reduced accuracy. To address this challenge, the patent describes a method of inputting the computational conditions into the model as “label information,” thereby enabling the model to be trained using a mixture of data generated under different conditions.(Summary based on Claim 1 of International Publication WO2022/260178. The wording of the claims at the time of registration may differ due to amendments made during the examination process.)This technology supports “general-purpose” atomic simulations that generalize to both crystals (periodic boundary conditions) and molecules (free boundary conditions). It reflects an intellectual property strategy characteristic of an AI company—distinct from that of materials manufacturers—by patenting the AI model training method itself.
In addition, Yokohama Rubber’s two patents are structured to cover the “forward problem (formulation → prediction of physical properties)” and the “inverse problem (target physical properties → formulation search)” in separate patents, while Fujitsu’s two patents combine the classical combinatorial optimization problem of similar structure search with machine learning and annealing algorithms.From a practical standpoint, it is interesting to note that the layers of patent protection—data preprocessing, model training, search algorithms, and operational systems—differ depending on each company’s business structure.
Looking overseas, AI companies, electronics manufacturers, and specialized MI vendors are each pursuing patent protection from their respective perspectives. The following are also registered patents that have been verified on the official patent database.
| Patent Number | Patentee | Key Technical Points |
|---|---|---|
| US 12,190,236 | DeepMind Technologies | Predicting the properties of new materials using embeddings of known material structures (Registered January 2025) |
| US 11,537,898 | Samsung Electronics | Reverse design that uses a GAN to learn the joint distribution of structure and properties and directly generates material structures from target properties (European corresponding patent EP3800586B1 also registered) |
| US 12,135,927 | IBM | “Expert-in-the-Loop” type search that learns experts’ acceptance or rejection judgments regarding AI-generated material candidates and reflects them in the candidate ranking |
| US 11,901,045 | IBM | A framework that learns features from a chemical database and generates new candidates by combining the features of existing materials |
| US 11,004,037 | Citrine Informatics | Uses machine learning to generate a capability map showing “achievable property combinations and their difficulty levels,” linking product design and materials development |
| US 10,984,145 | Citrine Informatics | Formulation design support that ranks candidate formulation recipes—including those with new raw materials—based on the likelihood of meeting target properties and constraints |
Claim 1 of DeepMind’s U.S. Patent No. 12,190,236 (priority date April 24, 2020; issued January 7, 2025) begins as follows:
"A computer-implemented method for predicting one or more properties of a material, the method comprising: maintaining data specifying a set of known materials, each having a respective known physical structure..."
(A computer-implemented method for predicting one or more properties of a material by maintaining data specifying a set of known materials, each having a known physical structure—followed by a description of the structure for property prediction using the identification and embedding of similar known materials) Source: Google Patents (US12190236B2)
The inventors listed on this patent include Tian Xie and James Kirkpatrick. This list overlaps with the authors of the GNoME and MatterGen papers, suggesting that patent applications for the underlying technology were steadily progressing behind the scenes of these much-discussed research findings.On the other hand, no patent application corresponding directly to Microsoft’s “MatterGen” itself was identified in this investigation (based on publicly available information as of July 2026). Please note that, as a general rule, patent applications are not made public for 18 months after filing; therefore, the fact that an application cannot be found does not necessarily mean that one has not been filed.
Practical Point: As seen with Samsung’s reverse-engineering patent, the basic architecture for material generation using GANs and generative models has already been patented in applications filed around 2020. Companies developing or implementing in-house material discovery systems using generative AI are increasingly needing to verify freedom-to-operate (FTO) with respect to the discovery methods themselves to ensure they do not infringe on third-party patents.
The greatest hurdle in patenting the results of MI is not inventive step or the status of an invention, but rather the disclosure requirements (enability and support requirements).The Japan Patent Office has clarified this point in its “Case Studies on Patent Examination of AI-Related Technologies” (published in three installments in March Heisei 29, January Heisei 31, and March Reiwa 6; a total of 25 cases, included in the appendix to the Patent and Utility Model Examination Handbook). The case studies set forth the following guidelines:
“An invention relating to an object presumed by AI to possess a certain function does not satisfy the requirements of enablement and support unless the detailed description of the invention includes examples in which the object was actually manufactured and the function was evaluated; this is the case unless the results of the AI estimation can serve as a substitute for the evaluation of an actually manufactured object.(Case 51, Case 52)”
— From the summary at the beginning of the Appendix to the Patent and Utility Model Examination Handbook, “Case Studies on Patent Examination of AI-Related Technologies”
Case 52, added in March Reiwa 6, deals precisely with MI reverse engineering (predicting chemical structures from luminescent properties). In this example, compounds A and B were obtained using AI; compound A was actually synthesized and its physical properties were measured, while compound B was based solely on prediction. The conclusion is clear.① Claims limited to Compound A, for which physical properties have been measured, satisfy the written description requirement; ② Claims for Compound B, based solely on predictions, and broad claims defined solely by property values violate the requirements for enablement and support; ③ Submitting a certificate of experimental results after filing does not cure the deficiency in written description.In other words, under current examination practice, a substance patent cannot be obtained at the stage of a “list of promising candidates generated by AI”; patent rights can only be secured within the scope that has been experimentally verified (as an exception, predictions may be considered a substitute for actual measurements if, for example, the accuracy of the AI’s predictions is verified within the specification itself).
Regarding claims for prediction methods, Case 50 (allergy incidence rate prediction) serves as a useful reference.Broad claims for prediction methods that do not specify input data (descriptors) violate the requirement for support unless the correlation between the data can be inferred from common technical knowledge; claims limited to verified combinations of descriptors were deemed valid. Furthermore, regarding inventive step, Cases 33 and 34show that inventive step is not recognized if the method merely “systematizes human tasks using AI” or “simply replaces a regression model with a neural network”; however, Case 34 demonstrates that inventive step is recognized when a significant effect is achieved by adding input data for which the correlation was not part of common technical knowledge.The practical guideline derived from these examination cases is that the patentability of MI inventions lies not so much in the novelty of the algorithm as in the “non-trivial selection of features and descriptors based on knowledge in the field of materials.”
Practical Point: When claiming with numerical limitations or parameters, it is necessary to include representative examples that define the scope of the claims, taking into account the support requirement established by the Grand Bench ruling of the Intellectual Property High Court (November 11, Heisei 17; Polarizing Film Case).The basic approach to specification drafting in the MI era involves designing the filing timeline as follows: “AI prediction → physical synthesis and measurement of a small number of promising candidates → filing a patent application for substances or compositions within the verified scope → filing an improved application after additional verification.”
It is said that the source of MI’s competitiveness lies not so much in the model itself as in the long-standing experimental database and the descriptors (features) designed from it.While these assets are well-suited for confidentiality because they cannot be reverse-engineered from products, filing a patent application generally results in publication after 18 months. Additionally, method claims present enforcement weaknesses, as it is difficult to detect their implementation on third-party servers (this paragraph provides a general overview of current practice).From a regulatory perspective, the March 31, Reiwa 7 revision of the Ministry of Economy, Trade and Industry’s Guidelines for Trade Secret Management clarified that even a combination of publicly known information may be recognized as non-public if acquiring it as data for AI training requires a significant amount of time and expense, thereby establishing a framework for protecting MI’s experimental datasets as trade secrets.If you choose to keep the information confidential, securing evidence (such as timestamps) to establish a right of prior use (Article 79 of the Patent Act) in preparation for subsequent applications by other companies is a standard practice.
Regarding patent applications naming AI as the inventor (the DABUS case), both the Tokyo District Court ruling on May 16, Reiwa 6, and the Intellectual Property High Court ruling on January 30, Reiwa 7, determined that the inventor under the Patent Act must be a natural person (the Intellectual Property High Court also referred to the need for legislative discussion).The United States (Thaler v. Vidal), the UK Supreme Court, and the European Patent Office have reached the same conclusion. The practical implication in the context of machine intelligence (MI) is straightforward: the natural person who creatively participated in defining the problem, designing the descriptors, selecting candidates, and conducting validation experiments must be identified and listed as the inventor.The Japan Patent Office is also examining, from an institutional perspective, how to protect “inventions involving reduced human involvement,” using materials screening as an example (Document from the Subcommittee on the Patent System of the Industrial Structure Council, March Reiwa 7).
① MI-related patents have been increasing since 2015, and in Japan, companies such as Resonac, Yokohama Rubber, Fujitsu, and PFN have secured registered patents across the layers of search methods, operational systems, and training technologies.② Overseas, companies such as DeepMind, Samsung, IBM, and Citrine have secured rights to foundational configurations, and registered patents already exist for reverse engineering using generative models.③ It is standard examination practice in Japan that substance patents are not granted for “materials predicted solely by AI” (Cases 51 and 52), and the design of proof-of-concept experiments and the timing of filing determine the success or failure of securing rights—these three points form the backbone of this article.
Pre-Filing Checklist for MI×AI Inventions
We continue to cover AI-related intellectual property trends on this blog. For insights on how to analyze pending patent disputes, please see our article analyzing the Kioxia v. Viasat patent litigation; for practical guidance on forming patent families, please also refer to our article explaining divisional application strategies.
Companies considering how to balance patent applications for MI and AI-related inventions (including the design of examples in the specification and the scope of disclosure of descriptors), prior art searches, FTO searches, and trade secret management are welcome to contact us via the inquiry form on the EVORIX Intellectual Property Office website.
[Disclaimer] This article is a general explanation based on publicly available information as of July 18, 2026 (including Google Patents, J-PlatPat, WIPO PATENTSCOPE, materials published by the Japan Patent Office, and press releases from various companies), and does not constitute legal advice.While the bibliographic information for all patents mentioned in this article has been verified on the actual pages of public databases, this does not constitute an interpretation of the scope of protection or an assessment of validity, nor does it indicate any business relationship between the companies mentioned and our firm.The claims of Patent No. 7403032 are cited as they appeared at the time of international publication; the wording may differ due to amendments made at the time of registration. Statistical values depend on the timing and definitions used in the source’s compilation. (Supervised by a patent attorney)