Edit/Keich
On March 23, at the Heart of Machines AI Technology Annual Conference, Wang Xuanze, founder and CEO of Creative Materials, shared the relevant content with the theme of "AI + Metal Materials: More Suitable for the Direction of Industry Landing" at the "AI x Science Forum".
The sharing discussed the problems that may be encountered in the process of landing in the AI industry, including the accuracy trap; the industry barrier is not technology, the transformation of traditional large factories is self-research; some problems in some toB fields; the nature of the black box is serious, customers do not recognize, and cannot replace key position personnel in the short term; the algorithm effect is amazing but the landing is difficult.
Wang Xuanze said that the industrialization of AI-enabled metal materials can effectively avoid or solve the difficult problems in the above AI landing. On the other hand, high-end metal materials are an often overlooked market, and with industrial upgrading and strategic transformation, the demand for localization substitution is rapidly amplified. The main difficulty in the high-end metal field lies in the long R&D cycle and excessive R&D investment, so the research and development of new materials using AI to empower has become the optimal solution for overtaking in curves.
https://www.bilibili.com/video/BV1SL4y177nR?t=7.0
The following is the content of Wang Xuanze's speech at the Machine Heart AI Technology Annual Conference, and ScienceAI has edited and sorted out the original meaning:
Good afternoon, everyone! Thank you very much for the invitation, and I am very pleased to be named to the Machine Heart 2021 "Top 10 Most Promising ScienceAI Startups" list. First of all, to briefly introduce myself, I was sent to Shanghai Jiaotong University through the computer physics competition, where I read a master's degree and majored in artificial intelligence. After graduation, I went to the first share of AI on the science and technology board, mainly doing artificial intelligence algorithms in the field of images, helping domestic and foreign head manufacturers to do camera built-in image algorithms. After leaving his job, he founded Minshi Technology, mainly doing industrial testing projects, and achieved some small achievements. Later, I participated in an AI+ small molecule drug startup project, and the specific algorithm framework was developed by me. Now it is committed to the research and development of AI+ basic materials. I am an entrepreneur and a relatively senior algorithm engineer, so I have seen some AI landing problems in the process of working on the front line for many years in the process of these AI landing projects. Today's sharing is also to share with you the problems I have encountered or seen, and why we finally chose the field of AI+ metal materials for in-depth research.
Problems that may be encountered during ai landing
I don't know if you have encountered the following problems in the process of AI landing.
First, the precision trap. Because the algorithm develops very quickly, there are some areas where the investment of the algorithm company and the improvement of the effect of the algorithm are not actually a one-time function relationship, somewhat similar to the situation of the sigmod function. After the algorithm has developed to a certain extent, it is necessary to invest a very huge amount of time, manpower, including the cost of energy to improve its performance. The difficulty of these algorithms to break through the upper limit is constantly increasing, and at the same time, due to the increasing degree of open source of the algorithm, the threshold of the algorithm is decreasing, resulting in some independent algorithm workers can also use the open source network to do some projects. After some small algorithm companies do some tuning parameters and model optimization for the deliberate demand points, they can further improve the effect of the algorithm. However, if you want to improve the effect or pursue the ultimate effect, you need to invest a lot of costs, and the cost of input is a geometric increase. And after investing a very large cost, the effect is only increased from 95% to 97%, and many end users cannot obviously feel the improvement of the algorithm, resulting in a good algorithm company, and its algorithm products will not be more expensive than the peers or the average in the industry. As a result, these top algorithm companies are quite good when they calculate gross profits, and once they calculate the net profit data, they are very bad and cannot make a profit for many years.
Second, the technical barriers brought about by algorithms are constantly compressing. Because its lower limit is increasing, it is more difficult to break through the upper limit, although the range is getting narrower and narrower. In some areas, on the contrary, the traditional barriers are still some barriers in this field, such as supply chain, channels, brands, costs, etc., and these barriers are still the most core barriers in this field. At the same time, on the basis of these barriers, some large factories in the field are also doing transformation, applying new technologies and AI technologies to empower themselves, and doing their industrialization upgrades, which will be more difficult for artificial intelligence startups. I have also encountered this problem in the process of doing image algorithms before. Now the image of the better mobile phone manufacturers have some self-developed teams, a long time ago we will feel that these teams are vulnerable, but more and more found that their growth rate is getting faster and faster, and the competitiveness is also getting bigger and bigger, which may also be a problem in the industry.
Third, for certain to B areas. It feels like any business model has some problems, such as a platform company, and its bigger problem is that it is difficult to achieve extremes when it wants to achieve the point of each market segment. Some platform-based companies use a general-purpose detection network, the scene points that can be covered are very limited, even if they cover this scene point, if a deliberate network is established in a special scene, it is adjusted and adjusted in details, the performance is far higher than the universal platform, the general network performance. This can be a problem in some areas of platform companies.
The bigger problem with service companies is that their growth model is not as good as that of platform companies or product companies, which may be similar to linear growth. Although the core algorithm in the order received is relatively similar, in most fields, the difference between each order is still very large. If a service company wants to expand its scale, expand its sales, or have a larger market share, it has to take a lot of orders and invest a lot of costs. Once the follow-up order cannot be received, there may be some operational pressure on the company's entire development. At the same time, these service companies, especially some customers in the domestic industrial field, it is difficult to provide their most core technology, core data and needs to these service companies, after all, it is the core secret, which will lead to service companies can never accumulate very core data and products. I've heard that some entrepreneurs have an idea, starting with a service company, taking some orders, getting demand from orders, making a product, and then doing sales. It sounds wonderful, but in fact there will be a lot of difficulties.
Now many companies, themselves in the industrialization of the transformation, a bit similar to their own use of technology to create a shovel, and then their own use of the shovel to mine, this is better than the platform and service companies, but will also face some challenges. The key point is that the core differences in each area lead to more challenges. In the process of industrialization transformation, choosing some better areas is a more critical thing.
Fourth, the nature of the black box is serious, the customer does not recognize, in the short term can not replace the key positions of personnel. Artificial intelligence, especially deep learning network black box nature is more serious. Many customers are not very accepting and not very recognized. Especially in safety-related areas, such as chemical safety. Once chemical safety produces safety accidents, it is very serious, and even a safety accident will cause billions of chemical companies to disappear. Chemical safety should be implemented to people, accountability to people, if a black box form of artificial intelligence network instead of the entire expert system or human evaluation standards, or a set of logical evaluation standards, there will be problems. Eventually, once a problem occurs, the accountability is given to the person, and it is found that it is a black-box nature of the network, and no one knows how to get the conclusion, which is completely unacceptable.
So, what do most fields do? On the basis of the original set of human systems, and then build a redundant set of neural network-based systems, which will also bring some problems. For example, it does not save labor, does not reduce costs, but increases costs because the system needs to be rebuilt. At the same time, in some other fields such as intelligent driving, telemedicine and other fields, even if our safety factor reaches the usable level, the cost of educating customers is very expensive. Smart driving, for example, has been advertised for years, and even in a few years these drivers will let go of their grips, free themselves from the steering wheel, and sleep in the cab.
Fifth, the algorithm effect is amazing, but it is difficult to land. For example, federated learning, it is very difficult to build an ecological environment across scenarios, such as the cognitive intelligence of the fire some time ago, and it is difficult to obtain some high-quality labeled data. Data is also a difficult problem for artificial intelligence to land in some fields. For example, industrial testing. Some steel plate defects, the production line is constantly running, can only produce a few or even a dozen defects a year, how to collect its data, it is difficult. We can only judge it through a semi-supervised way, or the logic-based recognition ability of traditional algorithms, but this is certainly not as good as the way it is trained with big data, and the performance and effect will vary greatly. And DeepFake, it's hard to imagine how compliant it is.
AI+ metal materials are more suitable for the application scenarios of artificial intelligence landing
If you encounter these problems, you may wish to turn your attention to the field of AI+ metal materials. Why do I think AI metal materials are more suitable for the scene of artificial intelligence landing? Because it has some characteristics, the most important feature is that its research and development cycle is very long, but the verification cycle is short. The research and development of traditional metal materials often has a cycle of several years, more than ten years, and decades, similar to foreign high-end metal material companies are old enterprises in World War I and World War II, and may have precipitated R& D technology after decades and hundreds of years of accumulation. Once the performance of the metal material meets the standard, it can quickly enter the production and sales stage, which can bring relatively large profits to the company. Therefore, the core pain point of metal materials lies in research and development, and the role that artificial intelligence can play in it is that the technology of artificial intelligence has a subversive new model in the entire metal research and development stage, which solves the difficult problem of long-term research and development, once the problem of research and development is solved with AI+ metal materials, some of the subsequent problems are actually relatively simple.
There are also many cases in China. For example, many companies are independent of the original technology to make a company, this enterprise does its own production and sales, so as to go public, and achieve better results, such as steel research Gaona, Western Superconductivity and Polarite. So this path is very feasible, at the same time, we ultimately provide the end user with not a set of algorithms and solutions, but the final product, parts, or a material, which can effectively solve the problem of artificial intelligence black box, customer disapproval.
In recent years, some new preparation technologies and advanced metal technologies will also make it more convenient to land applications. For example, 3D printing technology, it can be said that there is no need to form a factory assembly line of hundreds of people, dozens of people are responsible for dozens of 3D printing equipment can achieve billions of output value, and there will be no emissions, pollution problems, this is a better landing scene.
In terms of data, the accumulation of data in recent years can already support us to accelerate the research and development of artificial intelligence. Mainly since 2011, China and the United States have proposed material genetic engineering, which may not have been available at that time, but everyone has realized that data is very valuable, and both enterprises and research institutes are constantly accumulating data.
In the past one or two years, the high-throughput metal preparation technology has developed rapidly and has also entered a mature stage, which can make metal acquisition easier and cheaper, enough to support artificial intelligence technology to make certain breakthroughs in the field of metals.
We will set up a high-throughput laboratory in the first half of this year, and then we will continue to produce high-quality metal data and do our own research and development.
At the market level, there is also a very large market demand. The metal field is a market that is often overlooked, when it comes to the metal field, everyone will feel that this is low-end production capacity, cabbage prices, low profit margins, control costs, compressed costs, etc., indeed for some low-end metal materials. However, many high-end metal materials are still very tightly stuck by foreign card necks, and the high-end metal material market is very huge, more than ten trillion US dollars in the market, and China will also import a large number of high-end metal materials every year, because the domestic preparation cannot be made. At the same time, the current demand for localization substitution is also very strong, the government in some core areas even clearly require that there must be a localization substitution rate, and it is improved year by year, in the government's policies and national planning can clearly feel the policy in this regard. For example, in 2021, the Ministry of Science and Technology's major project of 6+1, that "1" is material informatics, that is, artificial intelligence + metal materials, which may be the demand side.
For the supply side, the most core technologies of AI+ metal materials have just entered the mature stage during this period of time, including artificial intelligence technology, metal high-throughput experimental preparation technology, 3D printing technology, etc., which have just reached the technical high point and are at the stage of potential energy vertex transformation to industrialization.
In terms of talents, this year's first batch of domestic material informatics doctoral graduates, about two or three years will have hundreds of doctoral masters released, and now AI + metal materials are now the hottest research direction in the field of materials. So, this is the best time for AI+ metal materials.
Creative materials for further study, AI empowers the research and development of new metal materials
The following is an introduction to how artificial intelligence empowers the research and development of new materials, and how to do further research in creative materials.
Generally speaking, the research and development mode of metal materials is divided into four stages. The first stage is the empirical discipline, the trial and error method. The second stage is to summarize some physical models and physical laws through a large number of experiments on the basis of the first stage, and guide our experiments. The third stage is after the rise of computers, put these physical models into the computer, and use the computer to iteratively simulate the calculation to further improve the speed of research and development. The fourth stage, the fourth paradigm, is the material informatics that is now used to reveal the more essential things of the material through the way of big data. Why can the fourth stage do some things that the first three stages can't do, because the first three stages are simulations based on physical models. How did this physical model come about? It is through the accumulation of some scientists for a long time that an approximate physical model is abstracted, because the difference between metal materials and other materials is that metal materials want macroscopic properties, and some disturbances will inevitably be introduced in the preparation process. For example, two or three oxygen molecules are entered, including uneven distribution of temperature fields, which is inevitable, so these formulas cannot fit the physical phenomena that actually occur with high redundancy, so the error is very large, large enough to be unacceptable. The fourth stage fits the physical model and process through the ultra-high redundancy of the neural network, the error is much smaller, and it is already a practical stage.
Through the organization of the image to find the corresponding relationship between the composition and the process, of course, we are not only using artificial intelligence algorithms to do, we are a system engineering, we will also use the DFT algorithm, material calculation method, the calculation of more accurate eigentometric quantities as network input parameters input network, improve the effect, roughly such logic.
Materials have a huge impact on the shape of human society, and materials have been developed for thousands of years, but the areas we are actually exploring are very, very limited.
Take ternary alloys as an example. More than 90% of the alloys we explored are low-entropy alloys, and this element may account for more than 80 to 90% of the entire alloy, and all other elements may add up to less than 10%. The exploration of medium entropy and high entropy alloys is very limited, in 2004, the concept of high entropy alloys was only proposed, and in 2010 some scientists explored this field, so we have a wider range of unknown areas for us to explore. Why has high entropy alloys not been explored for so long? The reason is that the traditional way of exploring this high-entropy alloy range fluctuates very much, and the cost is much higher than that of low-entropy alloys. If high-entropy alloys are explored by traditional methods, it is estimated that many scientists will accumulate generations of experience before they can barely extract a relatively limited physical model. But now with the fourth paradigm of artificial intelligence to study, we only need to accumulate a large amount of data, and use artificial intelligence to find the correspondence between component process and performance. After finding the correspondence, scientists may be able to extract more advanced materials science principles through the results obtained by AI algorithms, but the development of materials science in the future is likely to become such a model.
My sharing ends here. Thanks again for the invitation from Machine Hearts, thank you!
In addition, we are currently recruiting talents, if you are interested, welcome to join us! We desperately need people in materials science and artificial intelligence.
Resumes or general exchanges are very welcome!
Artificial Intelligence × [ Biological Neuroscience Mathematics Physics Materials ]
"ScienceAI" focuses on the intersection and integration of artificial intelligence with other cutting-edge technologies and basic sciences.
Welcome to follow the stars and click Likes and Likes and Are Watching in the bottom right corner.