MIT Study Aligns With USPTO’s Inventorship Guidance for AI-Assisted Inventions in Material Science Industry

December 4, 2024

A working paper published on Nov. 6, 2024, by Massachusetts Institute of Technology PhD candidate Aidan Toner-Rogers suggests that the patent system may soon see a marked increase in patent application filings due to increased productivity from artificial intelligence (AI). However, the impact on patent issuance is not as clear.

Toner-Rogers studied AI’s impact on innovation in a materials science lab and determined that in a test case in the field of materials science, “AI automates 57% of ‘idea-generation’ tasks, reallocating researchers to the new task of evaluating” the ideas for candidate materials that the AI model produced. By using AI to assist in idea-generation, “researchers discover 44% more materials, resulting in a 39% increase in patent filings and a 17% rise in downstream product innovation.”

An earlier McGuireWoods legal alert that reviewed the U.S. Patent and Trademark Office’s (USPTO) Inventorship Guidance for AI-Assisted Inventions explained that although AI “is not itself able to qualify as an inventor or co-inventor” of a patent, “if [a] natural person would qualify as a co-inventor with the AI, the invention is patentable with the natural person treated as the sole inventor.”

This alert considers how the USPTO’s inventorship guidance aligns with the AI-assisted innovation described in the study (which was designed before the guidance was issued) and highlights practice tips for those who utilize AI in the innovation process and seek patent protections for the results.

The Toner-Rogers Study

The study was conducted between May 2020 and June 2024 at a large U.S. firm that specializes in materials science. The lab focuses on “healthcare, optics, and industrial manufacturing,” and 1,018 scientists participated.

The study utilized “graph neural networks (GNNs) trained on the structure and properties of existing materials.” The GNN represents materials “as multidimensional graphs of atoms and bonds, enabling [the GNN] to learn physical laws and encode large-scale properties.” The GNN “undergoes three stages of training to optimize its performance: pre-training on a large dataset of known structures, fine-tuning with application-specific material properties, and reinforcement learning using tests of model-suggested compounds.”

“Scientists begin the search process by defining a set of target features” for a new material. “These features vary significantly across materials, but often include both atomic and large-scale characteristics determined by the intended application.” After the scientists input the desired characteristics, “the [AI] model generates candidate compounds predicted to possess these properties.” In other words, the AI model “generates ‘recipes’ for novel compounds,” by “begin[ning] with a known structure, add[ing] noise, and then revers[ing] the process to create a new compound.”

Once the “recipes” are created, they are evaluated by the scientists. “While AI designs have the potential to rapidly advance discovery, materials proposed by even frontier models are often unstable or display other undesirable features. Consequently, it is essential to pair the technology with scientists who can evaluate, refine, and test candidate compounds.”

The scientists “focus on selecting which compounds to advance, often involving the analysis of simulations or predicting material characteristics based on domain knowledge.” The scientists rank candidate compounds based on predicted quality and “synthesize the most promising options” to “reveal candidate materials’ true quality” and “to evaluate their properties.” They “rul[e] out a large share of candidates that do not yield stable compounds … conduct tests to assess [a compound’s] properties at both the atomic and macro scales … [and] subject it to real-world conditions such as heat, pressure, or human interaction.” “Once researchers create a useful material, they integrate it into new product prototypes that are then developed, scaled, and commercialized.”

Application to Guidance Examples

The study’s AI-assisted scientists filed 39% more patent applications than expected without AI-assistance. The USPTO’s guidance would apply to these filings. Along with its guidance, the USPTO published two examples with different scenarios that modified the basic facts underlying the example. The two scenarios from example 2 provide a benchmark to consider the scientists’ work.[1]

Manipulation of AI Output

In scenario 1, a ready-to-use deep neural network (DNN)-based AI model is used to identify candidate drug compounds to target mutated androgen receptor (AR) protein. Scientist 1, Marisa, identified well-known datasets and a target amino acid sequence to input into the AI model. Using these inputs, the DNN generated a series of outputs, which were ranked by binding affinity. Marisa selected the top six candidate compounds for further experiments and characterizations. Naz, a research fellow, synthesized these six compounds and identified the most promising one. After discussion, Marissa and Naz identified potential modifications to address deficiencies in the lead candidate. Naz created an intermediary compound, and Marissa created a modified compound with better properties. The university filed a patent application on the modified drug compound.

According to the USPTO, Marisa and Naz are proper inventors. Marisa identified the problem to be solved, the desired properties to be achieved and the proper input for the AI model. Marisa also identified the top candidate compounds for experimentation, determined appropriate modifications to optimize desired properties and synthesized a modified compound. Naz performed the experiments on the top candidate compounds, characterized them, determined appropriate modifications and prepared intermediary compounds. Although some of these steps “could be characterized as simply identifying a problem or reducing the output of the [AI model] to practice,” the steps of synthesizing the AI-generated compounds, characterizing them, and modifying them constituted a “significant contribution.”

Applying the USPTO’s guidance to the scientists described in the Toner-Rogers study yields parallels. The scientists identified the problem to be solved (the “intended application”), the desired properties to be achieved (the “set of target features”) and the proper input for the AI model (“a large dataset of known structures, … application-specific material properties, and … tests of model-suggested compounds”). The scientists also identified top candidate compounds for experimentation (“select[] which compounds to advance”), performed the experiments (“test”), determined appropriate modifications to optimize desired properties (“evaluate”) and synthesized modified compounds (“refine”).

Thus, it appears likely that the study scientists’ contributions are significant, like those of Marisa and Naz, such that they could be listed as inventors on patent applications resulting from the inventions.

Practice Note: When employing a ready-to-use AI model, merely identifying a problem to be solved by the model is insufficient. In this example, the guidance credits the manipulation of the AI output in some way, e.g., through prioritization, evaluation, synthesis and modification to best achieve the intended purpose. The study describes this as the exercise of judgment — taking the output of the AI model and applying the scientists’ skill and expertise to refine that output into a viable candidate compound. It is precisely this human effort, judgment and skill that the USPTO will continue to reward through patent issuance.

The USPTO does not provide an example in which the AI-generated output needs no modification and is immediately viable. But, given the discussion of example 2, it appears unlikely that mere synthesis of an AI-generated compound and integration into a new prototype, without refinement, would constitute a “significant contribution.” 

Specialization of an AI Model

In scenario 2 of the guidance, a DNN-based AI model is developed to identify candidate drug compounds having a sufficient binding affinity and other desirable properties. Scientist 1, Raghu, developed a DNN that accepted a coded representation of a drug compound and output optimized candidate drug compounds. Scientist 2, Marisa, identified and quantified desirable properties and defined a scalar objective function based on their optimum ranges. Raghu trained the AI model with existing datasets and fine-tuned it with observations of desired properties and AI-outputs. They input coded representations of six compounds into the AI model, and the model output six modified drug compounds, which Marisa synthesized in the lab. Marisa determined the most viable candidate compound, for which the scientists filed a patent application.

Raghu and Marisa made significant contributionsthey identified problems with existing drug compounds, set out to find a particular solution (i.e., a candidate drug compound with desired properties) and trained the AI model to that end with existing datasets and fine-tuning. “Without their contributions to … [the AI model], [the compound] could not have been created.” Moreover, Marisa recognized the modified compound’s value and synthesized it.

Applying the USPTO’s guidance to the scientists described in the Toner-Rogers study appears to yield similar results. The study scientists identified an “intended application,” set out to find a particular solution (i.e., a candidate compound with a desired “set of target features”) and trained the AI model to that end, with existing datasets and fine-tuning. But for the scientists’ work on the AI model, the compound could not have been created. The study scientists recognized the value of the most promising output and synthesized it, as Marisa did. This is where the most highly skilled scientists excelledin evaluating model-suggested candidate compounds and synthesizing the most promising options.

Practice Note: The USPTO example states that “[t]he natural person(s) who designs, builds, or trains an AI system in view of a specific problem to elicit a particular solution could be an inventor, where the designing, building, or training of the AI system is a significant contribution to the invention created with the AI system.” However, the USPTO cautions in example 1, scenarios 1 and 5: “[m]erely recognizing a problem or having a general goal or research plan to pursue does not rise to the level of conception” and “a person simply owning or overseeing an AI system that is used in the creation of the invention … does not make that person an inventor.”

When developing a specialized AI model, it is imperative that a natural person design, develop and train the AI tool to solve a particular problem (e.g., an intended application) in a particular way (e.g., through generation of an output with certain desired properties). The identification of desired properties appears to be key, as that connects back to both the problem to be solved and the way the AI tool will solve it. Importantly, it is essential that scientists properly document all steps undertaken to specialize the AI model, to demonstrate their contributions to its design.


1. Certain assumptions have been made. For example, the working paper states that the scientists “evaluate, refine, and test candidate compounds.” “Refine” is assumed to mean that the scientists modify the AI-generated compounds in some way. The working paper also states that scientists “defin[e] a set of target features … determined by the intended application.” “Intended application” is assumed to mean the problem to be solved.

Subscribe