A World Where Data Dictates the Development of New Materials
- Date : 19-12-31
- Views : 234
I recently switched out my outdated smartphone for a new one. I ask the personal assistant program that came with my phone about the weather every morning and request playlist recommendations. I’ve only been using this phone a month, but the imbedded artificial intelligence already has a good understanding of my everyday habits - what genre of music I like, what apps I most frequently use, when I come home, and much more.
However, this doesn’t cover the extent of the capabilities of artificial intelligence. When grand master Go player, Lee Sedol, was defeated by Google DeepMind’s AlphaGo three years ago, people were shocked not just because a machine beat a person, but because human arithmetic skills were already no match for that of machines.
What was even more astounding was that machines could be good at “creative thinking,” something that was considered the essence of human brain activity. This was shown when AlphaGo found moves that were more advantageous by studying (machine learning) prior information (data) created by humans. AlphaGo has since retired from the game of Go, but DeepMind has developed a new program called AlphaFold and commenced research on the protein structure of living organisms, with the grand vision of curing incurable diseases.
Efforts have been made across the board to generate creative results with machine learning algorithms. From a material scientist’s point of view, artificial intelligence can potentially develop cutting-edge new materials. Let’s say that material researchers across the world pool knowledge to create a database required for AI machine learning. Like the goose that laid a golden egg, AI may give birth to new materials, one after another.
Nonetheless, data related to new materials development is still lacking in both quantity and quality. Much like how people gain insights from experience, AI comes up with better options by studying large amounts of data. This is why quality data is required. Quality big data is easily generated in healthcare, marketing, and other industries, thus making the application of machine learning in these areas particularly useful and efficient. Yet so far, the same quality and quantity of information is not available to developers of new materials.
The difficulty in accumulating information about different substances makes data accumulation in the materials field challenging. Using machine learning with data sets about different materials is meaningless. Data needs to be clustered for each specific material to produce solutions.
We need to apply big data categorized for each different material to AI. However, there are only a few researchers that study specific substances, which means that acquiring big data in a short timeframe would be challenging even if we were to pool all the information that is out there. We need to go a step beyond research data acquisition and accumulation and strive to artificially generate the data that is needed.
Conditions such as synthesis temperature, substance, composition ratio, and other independent variables used to manufacture different materials are required as input values to generate relevant data. Electricity and heat conductivity, solidity, and other values become the output. Each individual researcher has different purposes in mind, which is why the input value is adjusted and controlled in different ways. Ample data would produce statistically meaningful results even if the study were to be randomized. If there isn’t sufficient data, a top-down approach can be used, where variables are controlled for specific purposes, and each individual researcher provides specific input values to accumulate controlled data.
A measurement and analysis platform catered to the nature of materials should be created to make data accumulation more efficient and increase the reliability of output values. Currently data is generated in Korea using numerous analysis tools. Data hasn’t been accumulated because each individual researcher requests measurements, and analysis is also provided on an individual basis.
Even if researchers agree to accumulate analyzed data, the data will be useless if we fail to collect and manage information about what analytic sample and independent variables were used to generate that data. A measurement and analysis platform built on the basis of features of specific materials through which material analysis data is processed would make it possible to efficiently manage data clusters.
Data could be generated for materials where the underlying technology has reached a sophisticated level of maturity. This could also enhance the effectiveness of machine learning. It would be wise to build a measurement and analysis platform and data clusters for promising materials that meet these conditions. A dedicated commitment to gather the necessary data will facilitate the development of new materials with the assistance of artificial intelligence.