A New Approach to Materials Science Development: Will Artificial Intelligence replace materials scientists?
How AI as a tool is helping us discover new materials faster?
The idea comes from Edward O. Pyzer-Knapp, a senior engineer at IBM's research and development center in Europe who has long studied the power of artificial intelligence to drive scientific discoveries, and whose 2022 book "Deep Learning for Physicists" drew academic attention.
In collaboration with the Shanghai Ceramic Institute of the Chinese Academy of Sciences, Knapp conducted a systematic study of AI's impact on materials development, compiling a report published in 2022 in Computational Materials, a journal of Nature. The report is lengthy and highly professional, with a particularly insightful insight into the entire field.
Painful Points in the Materials’ World
To say that artificial intelligence as a new means to facilitate the development of new materials, we first need to know, what is the problem of the materials industry now? Why does AI have to be introduced to solve this problem?
The core painful point is that there are so many kinds of matter that the world of materials is stuck in a muddle. The pace of development of new materials in the last 100 years has been astonishing, so fast that you may not even realize it, for example, plastic, a material we can't live without today, only became popular in the 1970s. Since then, new types of plastics have been invented, and we come into contact with more than a dozen different types of plastics in our daily lives.
But too much of a good thing, as the saying goes, is that materials science is entering a "more is less" phase.
As we all know, matter is made up of atoms, and different atoms, arranged in different ways, can produce different substances. Based on this rule, it is possible to estimate the possible types of matter on the order of 10 to the power of 108. That's a huge astronomical number, because the total number of atoms in the entire universe is only about 10 to the power of 80. To put it another way, when we develop new materials, it's like picking up seashells on the beach, and there are a lot of them. When we first start picking up any shell, it feels fresh and we pick it up quickly. But by the time the shell bags are almost full, the numerous shells on the beach become a nuisance, for it is too much work to keep sifting through to find one that looks better than the one before.
In the past, in the development of materials, we also relied on some routine to find new possibilities. However, these patterns are often just empirical rules. Whether there will be anti-common-sense routines appear, for human beings, this is a huge amount of work.
But if AI uses certain algorithms, it will be possible to achieve new breakthroughs. The algorithm for screening new materials also relies on big data input, and AI conducts deep learning and retrieval analysis to see which structures are most likely to work.
Through this learning, the machine can train new models and screen them out to a database, where the new material is tested to determine the best target after simulating real-world conditions. It could even give researchers instructions on synthesis methods to try out. In this way, the search for new materials is no longer a rambling affair, but can be narrowed down to find new targets faster based on the results given by artificial intelligence.
Practical Applications
Ok, so this algorithmic logic sounds good, but how does it work in practice?
The researchers used the algorithm to find a new class of materials called PAGs. PAG, which stands for photoacid-generating agent, is one of the components of photoresist. When exposed to light, it produces an acidic substance that changes the properties of the photoresist and is used in chip processing. PAG is such a cutting-edge material that it cannot be studied carelessly, and it can be expensive. You know, as a researcher, every time you develop a PAG, you have to use it on an actual lithography machine before you know what it's like. It is not uncommon for a project to develop a series of hundreds of potential materials, using conventional screening methods. However, very few of these are likely to be used or even eliminated altogether. However, according to the logic mentioned above, the use of artificial intelligence in the development of new materials will be very clear, save money and effort.
Among the known PAGs, the materials already in widespread use are not perfect, and they pose risks of bioaccumulation and toxicity. Simply put, if you are exposed to these PAGs over a long period of time, the body may accumulate these components and become toxic. So can artificial intelligence be used to discover new healthy and environmentally friendly PAGs? That's true, here's how, in the first step, the researchers put more than 6,000 patents, papers, and data sources into the platform and let the machine do deep learning. It would take six months to a year for a researcher to read them in person, based on a typical length of work. Learning from these findings, the machine first sifted out the structures of about 5,000 PAGs. Based on this, some PAGs that do not meet the requirements for key indicators are eliminated, and the machine classifies them according to their parameters to determine which structures are conducive to the final goal. This allows the machine to generate a database of about 3,000 candidate targets based on what it learns.
These targets, most of which are entirely new substances, have not been studied before. The entire database takes about six hours to create. The same amount of work given to a regular scientific team would normally take several months.
From more than 3,000 candidates, the machine continues to learn, builds new models, and eventually selects about 400 of the most likely targets to run simulations on. What’s more, these simulations don't have to be performed on the lithography machine, and they're also screened for the best performance based on the machine learning results. In fact, by improving the algorithm, the filter doesn't even have to run through all the targets. The researchers tested just over 90 candidates before finding the most desirable one. The machine then recommended a one-step synthesis based on the results. “One-step" is a chemical term that means when ingredients are added to a reaction facility, the product can be made directly, with no additional work required. Materials made by the one-step process will have an advantage when it comes to industrialization. According to the algorithm, the researchers tested the scheme and did get the ideal results. In other words, the application of artificial intelligence to the development of new materials is highly feasible. In actual operation, there are still many indicators to consider. It is often the case that a material is excellent in one aspect and bad in another, which makes it difficult for the machine to evaluate it. This also answers the important question I mentioned at the beginning, that is, will materials scientists lose their jobs when artificial intelligence is applied?
According to the current logic of machine learning, the work of materials scientist is still irreplaceable. Not only do original scientific research data need to be provided for machine learning, but also some judgments on the properties of new materials cannot be completely left to machines to complete.
In any case, for the field of new materials, artificial intelligence has opened up a completely new mode, the future is looking forward to.