- Статьи
- Science and technology
- "Thanks to AI, the time for creating new materials will be reduced to two years"
"Thanks to AI, the time for creating new materials will be reduced to two years"
Russia has developed a neural network that can recognize the chemical formula of a material based on a photograph taken on a microscope. Valentin Ananikov, Academician of the Russian Academy of Sciences, Head of the Laboratory of the Zelinsky Institute of Organic Chemistry of the Russian Academy of Sciences, told Izvestia about this in an interview. He also said that at the next stage, based on neural networks, scientists will create digital reactor programs and methods for scaling chemical reactions from a test tube to a factory — this is how artificial intelligence will revolutionize domestic chemistry.
"Our country is facing the task of developing micro-tonnage chemistry as soon as possible"
— Valentin Pavlovich, tell us about the current developments of scientists at the Institute?
— The main directions are those related to improving people's lives and the development of the chemical industry. This is how the institute creates new catalysts, effective agrochemicals, medicines, energy materials, and much more. Promising research is the processing of plant biomass (for example, cellulose and lignin) into bioplastics. It is a material that is comparable in its properties to modern plastic, but it is easily regenerated and does not harm nature and humans.
At the same time, scientists are designing new technologies that accelerate the discovery of chemical reactions and help produce products faster and cheaper. Another important task for the country is the development of universal methods for scaling chemical processes, from obtaining a substance in a test tube to its production at a chemical plant.
— Can you give examples of significant developments?
— In the field of digital developments, scientists at the IOH RAS have created the world's first neural network, which determines its molecular formula based on a photograph of a substance taken on a microscope. The program analyzes with high accuracy and recognizes even substances that are close to each other.
The capabilities of this software package were first demonstrated using the example of special phosphonium compounds that are used as antiseptics. But most importantly, the possibility of using neural networks to solve problems that seemed fantastic until recently was shown.
Another project is the development of neural networks for the rapid detection of materials based on their spectrometry. This is a complex data set with many different signals. Their study requires a lengthy "manual" interpretation. We have created a software package that allows you to perform these operations quickly and automatically.
— What is artificial intelligence for chemistry?
— Artificial intelligence programs have rapidly burst into our world. Most of us use these apps on our gadgets, computers, and smart devices on a daily basis. They simplify everyday life. However, as far as scientific applications are concerned, artificial intelligence has been to some extent an end in itself.
Now there is a turning point. We are coming to the conclusion that these tools should be aimed at solving specific technical problems of humanity — in industry, economics, medicine, chemistry and other fields.
— What areas related to AI are being developed at your institute?
— The project "Digital Chemistry" has been implemented on the basis of the IOH RAS for the second year. This is a major program for the development of a new generation of chemistry. Substantial funding has been allocated for it. As part of the project, scientists are exploring how to use neural networks to predict new materials, synthesize them, and develop effective catalysts. And, most importantly, they are creating techniques that will help implement successful developments obtained at the laboratory level into full-fledged chemical production.
Our country is currently facing the task of developing micro-tonnage chemistry as soon as possible. It produces up to 100 tons of products per year. To create them, it is necessary in a short time to develop technologies for obtaining a large number of various chemicals in these marketable quantities.
"Thanks to AI algorithms, the catalyst design time is reduced by 2-3 times or more"
— What needs to be done for this?
— It is very difficult to accomplish this in a short time using traditional means, but, for example, thanks to artificial intelligence tools, it is possible to start developing effective catalysts. These are substances that accelerate a chemical reaction. Often, one such molecule contributes to the production of thousands and millions of other molecules.
In the chemical industry, 80% of all processes are catalytic. It is estimated that up to 35% of the world's total GDP is generated through the direct or indirect use of catalytic processes. However, their design is a complex and expensive process that, on average, takes from four to six years. And this is critical! Especially now, when humanity needs many new molecules to solve the challenges it faces. At our institute, scientists create and train neural networks to accelerate the development of catalysts.
— How does this happen?
— For example, using neural networks, chemical processes can be flexibly modeled, varying the conditions and components of the reaction, and the size of the catalyst particles. In addition, these programs can quickly learn from archived data and use it to develop new reactions. As a result, thanks to artificial intelligence algorithms, the catalyst design time is reduced by two to three times or more.
From a global perspective, we can say that the country that is the first to create neural networks for accurate prediction of catalytic processes will make a technological breakthrough and take a leading position in the chemical industry. And we strive for this.
— Can you give examples of how such neural networks work?
— In many reactions, metal nanoparticles act as catalysts. They are dynamic in nature and can change over the course of a reaction. In particular, several atoms can "break off" from the catalytic center and combine with each other, forming a complex mixture. In other words, the system is evolving, and you need to understand which components are most active in this process. Studying them using traditional methods will take millions of hours.
We solved the problem using neural networks. First, a person analyzes a small amount of data, and then the program, having learned, analyzes arrays of information in a similar way. Unlike a scientist, she performs procedures thousands of times faster, and at the same time does not get tired.
Using this technique, we have for the first time fully traced the movement of all the catalyst particles throughout the entire reaction. This approach is called "4D catalysis" because it involves studying the chemical process both in space and in time. This makes it possible to understand the main factors that accelerate reactions and create more efficient catalysts.
"The competence of a scientist is the generation of ideas and hypotheses"
— How can a chemical reaction be transferred from the laboratory to production?
— This is a really key issue right now. Scientists can produce any substance in the amount of several grams, but in order for synthesis to be of practical importance, tons are needed. However, scaling a chemical reaction is a critical area where many projects fail.
One of the solutions to the problem that scientists at the IOH RAS are working on is the creation of a digital copy of the reactor. All the nuances of the chemical process can be worked out on a virtual model. Then the version of the equipment tested in this way can be easily and quickly printed on a 3D printer. This dramatically changes the whole picture of chemical production.
Previously, the design, manufacture of parts and assembly of the reactor took up to several years. Now it can be done in a much shorter time, and the reactor can be printed on a 3D printer in a few days. This is especially true for micro-tonnage chemistry, where small-volume reactors are needed. They often fit entirely into a 3D printer's camera, and they can be printed in one cycle.
— How will new materials be created using AI in practice?
— Let's say that materials engineers need a new material with specified properties. They turn to chemists. Experts first predict its structure. (These techniques have already been well developed). The next step is to synthesize a small amount of the substance to test its properties. If necessary, the procedure is repeated. Then, with the help of a digital reactor, the technology of scaled production is being worked out. Neural network tools can be used at all these stages.
I would like to note that this is a complex task, which requires combining the efforts of different scientific and production teams. Such technological associations can be created nationwide. As a result, if the program is successfully implemented, AI will reduce the time required to create new materials to two years. For comparison, this process now takes eight to ten years.
— When will AI be able to create new materials without human intervention?
— There are no prerequisites for this now. Artificial intelligence is important as an assistant — for processing data and performing routine operations. Neural networks do not yet have practical mechanisms for conducting experiments and working directly in the laboratory — this is the prerogative of a scientist.
In other words, ideally, a person suggests the idea of an experiment and chooses a way to implement it, and assigns the program to work out the process. Next, the natural mind must control the actions of the "silicon" designer, and the person himself controls the experiments.
On a more global level, the competence of a scientist is to generate ideas and hypotheses, and then combine them into larger scientific constructions. The advantage of AI is in the thorough processing of all available information.
For example, in large-scale projects, researchers now manage to process only 10-20% of the experimental results. This is something that, by virtue of intuition and experience, seemed interesting to them. The rest remains without interpretation due to the high cost of time. Also, scientists do not use negative results. However, these are valuable materials, the so—called "dormant" data. Involving them in scientific research through the training of neural networks will significantly increase the effectiveness of research.
Переведено сервисом «Яндекс Переводчик»
