Artificial intelligence (AI) and machine learning (ML) are already key technologies of the 21st century and have found applications in diverse fields, from search engines and social media algorithms, to self-driving cars and medical diagnostic tools. AI also has the potential to revolutionise the development of new materials, which until now has been an empirical and relatively slow process. In this article, we take a look at the prospects for developments in this field, as well some of the pitfalls which will have to be overcome when it comes to patenting AI inventions in Europe.
The possible applications of AI and ML in materials science and engineering are myriad and have been set out in detail in various review papers.1, 2, 3, 4, 5 Example applications of AI include:
- data analysis (e.g. solving crystal structures from diffraction patterns)
- screening databases of known materials (e.g. identifying suitable materials for a particular application based on known properties)
- property prediction (e.g. predicting mechanical, electrical or optical properties for known or new materials)
- materials modelling (e.g. simulating the behaviour of materials under different sets of conditions)
- inverse design (e.g. designing new materials having tailored properties)
- selection of promising experiments (e.g. for investigating particular structure-property relationships)
- identification of synthesis routes (for existing or predicted materials)
- text mining (e.g. searching research papers for material structures, properties and synthesis routes)
A particularly ambitious goal is the development of entirely autonomous experimental systems which combine some or all of the above techniques with high-throughput experimental and/or computational methods to design, carry out and analyse experiments without human intervention.
Development of AI methods in any of these areas should lead to significant speed-ups in materials design, synthesis and characterisation.
For example, the interpretation of X-ray diffraction (XRD) patterns can be a laborious process and many complex patterns can be effectively insoluble for human experts, even when equipped with advanced computational tools. Researchers have, however, already developed a deep-learning AI method which can identify and quantify inorganic compounds in complex multiphase XRD patterns associated with materials useful as LED phosphors.6
Similarly, while state-of-the-art ab initio methods such as Density Functional Theory (DFT) enable highly accurate predictions of material properties at the atomic level, such calculations are slow and computationally taxing, restricting the time and length-scales which are accessible with current computers. AI models, however, can be trained on reference DFT calculations of small structures and then used to run much larger-scale simulations. In some cases, AI models can even be fit directly to experimental measurements and produce more accurate results.2
Research in this area is still at an early stage, particularly as concerns the development of entirely autonomous experimental systems. Nevertheless, it seems likely that the use of AI in materials science is set to grow rapidly. Now is therefore the perfect time to consider what types of patentable inventions may be made in the field and what challenges may arise.
In order to get an idea of the growth in patent filings concerning AI applications in materials science thus far, our IP Analytics team searched worldwide for patent applications published in the CPC classifications C01 to C14 (which cover “Chemistry” fields such as inorganic chemistry, glass, cements, ceramics and organic chemistry) and C21 to C30 (which cover “Metallurgy” fields including iron, ferrous and non-ferrous alloys and crystal growth technologies) and which include the words “artificial intelligence” or “machine learning” in the description. We identified a relatively steady growth in the number of applications published since 2011.
In many of the applications we uncovered, AI is used to identify existing or new chemicals, compositions or materials which have desirable properties. Of course, not all of these applications claim inventions which necessarily rely on artificial intelligence. In some cases, for example, AI is simply provided as a possible way of carrying out a particular step in a method (such as analysing data) which could also be carried out by other means. However, the results still indicate the growing use of AI methods in this field.
Patent applications relating to AI face two main hurdles (other than meeting the usual requirements for novelty, inventive step and industrial applicability) at the European Patent Office (EPO).
First, the EPO must be convinced that the AI-based invention defines “technical” features which solve a “technical problem” in a non-obvious manner; “non-technical” features that do not contribute to solving the technical problem are not taken into account. The EPO considers that the computational models and algorithms which underpin AI are per se of an abstract mathematical nature and thus non-technical. However, they may contribute to the technical character of the invention (i.e. contribute to solving the technical problem) by:
- application to a field of technology; and/or
- being adapted to a specific technical implementation.
Examples of “technical” applications of AI include the identification of a new chemical composition or the processing of an image to derive information. It is important to note, however, that generic applications such as “controlling a technical system” are unlikely to confer a technical character to AI inventions. Accordingly, inventions in which AI is applied to specific processes in the development of new materials are likely to overcome this first hurdle.
Second, the EPO will scrutinise whether the invention is sufficiently described. Article 83 of the European Patent Convention requires that the application discloses the invention “in a manner sufficiently clear and complete for it to be carried out by a person skilled in the art”. EPO Examiners must therefore consider whether there is enough information in a patent application to allow a person skilled in the art to reproduce the claimed invention, and this is where AI-based inventions in materials science are most likely to face challenges.
The requirement for sufficiency means that the application must include a description of at least one way of carrying out the invention. The application should also include sufficient evidence to make it at least plausible that any technical effect associated with the invention can be achieved across substantially the whole claim scope. When assessing sufficiency, the EPO takes into account the common general knowledge of the skilled person at the priority date and accepts that a limited amount of trial and error may be needed to put the invention into effect. However, the burden on the skilled person must not be undue and no inventive skill should be required to work the invention.
Applications in the chemical sciences and in AI face similar types of sufficiency problems because it is often unclear why inventions in either field work. For example, a chemist may not know why chemical composition A has better properties than chemical composition B, or why yield of a particular reaction product X increases when a reaction variable Y is changed. And for AI, an inventor may not know why a particularly strong weighting associated with a certain parameter in a model leads to better results. Since results in both fields can be unpredictable, it is often necessary to provide the skilled person with more information (as compared to mechanical applications) for the invention to be reproducible.
Therefore, in the context of purely chemical inventions, where the application claims a new chemical composition, a description of the composition per se is unlikely to be enough and it will usually be necessary to include full experimental details (including starting materials, apparatus and reaction conditions) of at least one way of making the claimed product. If the claims of the application are in some way limited by a particular technical effect (for example, where a method provides a particular result or a composition is used for a particular purpose), it is generally necessary to provide experimental data to prove that the result is actually obtained by the invention.
In the context of AI, especially in cases where AI forms the core of the invention and is not merely one way of implementing the invention, a sufficient description of the invention may require disclosing the principles underlying the data included in the dataset and any assumptions or parameters built into the AI model. Similarly, if the invention relies on the achievement of a particular result (such as better distinguishing between different types of image), data should be provided to prove this.
For applications in which AI is used in material science, it may be necessary to satisfy all of these requirements simultaneously, which could be challenging.
As an example, one application we identified in our search concerns an AI method for designing an alloy microstructure. A neural network is trained to correlate microstructural features with material properties using a training data set of micrographs taken from alloys having varied compositions and properties. The claimed method uses the neural network to identify a set of microstructural features capable of achieving a desired combination of properties for a particular use. The application also claims a step of manufacturing an alloy having the identified microstructure.
This example patent application has received two main sufficiency objections in Europe.
First, the Examiner has objected that there are no known generic theoretical models linking alloy properties to microstructural features and that such a model is not disclosed in the application. The Examiner also doubts that the AI model as disclosed can predict optimized microstructures for an alloy having desired properties (i.e. inverse design, as discussed above). Instead, the Examiner thinks that the AI could potentially be used to analyse an image of a previously unknown alloy structure and predict its properties (i.e. property prediction), but this is not what has been claimed. The Examiner has therefore objected that excessive experimentation would still be required for the skilled person to determine suitable structural features for a given set of properties.
Second, the EPO Examiner has objected that the application does not explain how to manufacture an alloy having a particular microstructure identified by the AI. The Examiner notes that developing a manufacturing method for a particular alloy requires extensive experimentation and that the skilled person would first need to consider whether any structure output by the AI was even chemically or physically feasible.
This case highlights the difficulties which can be encountered when prosecuting applications at the overlap between AI and materials science. It also reflects the problems faced by researchers in the field. For example, DeCost et al1 noted a lack of suitable mechanistic models in materials science which could be used in scientific AI to compensate for the relative paucity and expense of materials data compared to fields such as image recognition and natural language processing. Most research so far has also focused on exploring the unknown properties of materials known to exist, rather than developing entirely new materials.3 One reason for this is that we do not yet have predictive theories to tell us how to make a compound, given its structure, or even to determine whether a compound can exist or be made at all.2
The prospects for applications of AI in materials science are promising and it seems likely that AI-enabled materials design will contribute in some way to solving problems in energy storage, climate change, healthcare and communications technologies. However, research is in its infancy and there are many challenges to overcome, including the development of physically-meaningful heuristic models and the expansion of materials databases.
Patent offices are also only just starting to grapple with applications in this area, although at present we see sufficiency as being the major hurdle to patentability. We therefore recommend that any such applications are drafted with input from attorneys experienced in both AI and materials science, and, where possible, that experimental data is used to support both computational and materials aspects of the application.
Michael Ford is a Senior Associate in our Materials team. If you would like further advice on this topic, please get in touch.
1 Scientific AI in materials science: a path to a sustainable and scalable paradigm, BL DeCost et al, Machine Learning: Science and Technology, Volume 1, Number 3, 2020.
2 Artificial intelligence is aiding the search for energy materials, P. Patel and S. P. Ong, MRS Bulletin, Volume 44, Issue 3, 2019.
3 Using artificial intelligence to accelerate materials development, P. Ball, MRS Bulletin, Volume 44, Issue 5, 2019.
4 Autonomous experimentation systems for materials development: A community perspective, E. Stach et al, Matter, Volume 4, Issue 9, 2021.
5 A Review of the Application of Machine Learning and Data Mining Approaches in Continuum Materials Mechanics, F. E. Bock et al, Frontiers in Materials, 15 May 2019.
6 A deep-learning technique for phase identification in multiphase inorganic compounds using synthetic XRD powder patterns, J-W Lee et al, Nature Communications 11, Article number 86, 2020.