In 2007, Apple contracted with Corning Inc. to use their durable, thin, scratch resistant Gorilla Glass for the first iPhone. Since then, Gorilla Glass has come out with five new versions of the glass, each one thinner, more scratch resistant, damage resistant, and even cleaner than the ones before (1). The most recent version, Gorilla Glass 6, claims that it can survive drops from 1.6 meters on hard, rough surfaces (2). Apple is not the kind of customer you want to lose, but with new software updates every few months and a new iPhone once a year, the expectation of high speed innovation abounds. Corning Inc. is not the only company that faces pressures to innovate when it comes to the materials that make up their products. Companies spend millions of dollars and many years to create products with new material makeup. This process involves understanding many different variables that go in to the equation — the types of materials, how past materials performed, how one ingredient interacts with another, the desired properties of the final material, the specific use cases of the material, and the conditions in which it will be used, just to name a few. For humans to try to consume and derive insights from all of these variables requires years of training and high levels of expertise, and even then, the process of materials discovery requires extensive experimentations and even serendipity at times.
Recently, researchers and companies have started to use machine learning to reduce the time and cost of materials discovery (3). One company, in particular, is focused entirely on using AI to assist consumer packaged goods companies in product development. The company, Turing, which is backed by YC, claims to dramatically improve the speed and ROI for companies developing new materials. By using machine learning, Turing can reduce a process that usually takes 90 days to just 27 days, and can increase the ROI 14x doing so.
As data collection and production flourish, companies are better positioned to use machine learning to detect patterns and recommend new combinations of ingredients to create new products. Several successful use cases of machine learning in materials science include predicting new stable materials and speeding up the calculations of numerous material properties (3). New products for a company may mean more than just an operational edge, but a new source of revenue and a solution to problems yet solved. Machine learning is being used by researchers and companies for materials discovery in fields that impact the food we eat, the air we breath, and even the next chapters of innovations in healthcare.
As society looks for alternatives to the traditional meat industry, machine learning can be used to help predict the ideal combination of ingredients to produce meat-like meat alternatives. One company at the forefront of this move is NotCo. The company is using AI — charmingly called Giuseppe, named after an artist known for painting fruit — to predict ingredient combinations to create meat-like plant-based alternatives. As NotCo improves their data collection methods, the potential for data to drive their recipes will grow. Early on, the target for their model was animal-based flavors and textures. The kinds of data they used included: spectrometry data, texturometer, viscosity, pH, along with data collected from different scientific equipment that analyzes properties of food. The CTO of NoTCo reports, “You improve orders of magnitude because you get more data. And sometimes it's not just about more entries, but we also need new variables, new scientific machines that add new variables that supplement with better perspectives.” (7)
Direct air capture technologies are used to remove carbon from the air. The device pulls in air and filters out carbon with a CO2 sorbent, which is similar in concept to a sponge that absorbs gases. Researchers Andrew S. Ross and Evan D. Sherwin from Harvard University and Carnegie Mellon respectively wrote on the topic of material discovery for sorbent design,
“While CO2 sorbents are improving significantly, issues still remain with efficiency and degradation over time, offering potential (though still speculative) opportunities for ML. ML could be used to accelerate materials discovery and process engineering workflows to maximize sorbent reusability and CO2 uptake while minimizing the energy required for CO2 release.” (5)
Combining data on molecular structures and their behavior or properties, along with data regarding the laws of physics, materials scientists can create models that predict, given a certain molecule, what is the likely behavior or property of that molecule. These insights are valuable when determining what chemical ingredients to use for a certain type of material given the desired specificities of that material, in this case, the ability to absorb CO2 from the air.
Traditional drug discovery can take many years and millions of dollars. One promising advance in the field of drug discovery is the use of machine learning by scientists to reduce the time and cost that goes into discovery and production. In order to discover a new drug, there are many different relevant data sources to consider including — existing treatments, new data on a particular disease, molecular compounds and their effects, and information from new medical technologies, among many many others. Testing and domain expertise incorporate all of these data points and discover, then narrow in on possible options. Google parent company, Alphabet, has announced a new company earlier this month, Isomorphic Labs, to lead an AI-first approach to drug discovery. The company claims it will use machine learning to build models that can predict how drugs will interact with the body. Demis Hassabis, CEO of Isomorphic Labs wrote recently, “We believe that the foundational use of cutting edge computational and AI methods can help scientists take their work to the next level, and massively accelerate the drug discovery process. AI methods will increasingly be used not just for analyzing data, but to also build powerful predictive and generative models of complex biological phenomena.” (6)
Another example of machine learning for materials discovery is 3D printing materials. As the field of 3D printing expands from creating toys and trinkets to artificial organs and houses, different types of materials must be created — you can’t use plastic for an artificial organ as you do for a toy soldier.
In deriving these materials, again, there are extensive variables that go into the discovery process. Once a model is given data to train on, like data regarding different chemical ingredients and their expressed properties, new potential ingredients can be added as input and the model can predict the likely outcome. Some basic questions when imagining new materials may include, what type of chemical ingredients should I use? How do those chemicals interact with others? How do those chemicals interact with the production process? In the same way that humans would, a machine learning model can learn the relationship between inputs and outputs; This gives the model the capability to identify whether a material is, for example, likely to be flexible and strong. One team from MIT used machine learning to create a 3D printing ink that hardens under UV light (4). In order to do this, they selected certain chemical ingredients, fed that data into the machine learning model along with information about the desired properties (toughness, stiffness and strength) and the model produced recommendations on the different levels of chemicals that would result in the desired properties. For more information on breakthroughs in 3D printing with machine learning, click here.
Machine learning continues to demonstrate the potential to discover new compounds, combinations and products. ML models are often used to impact the way we interact with the virtual world, like Netflix’s recommendations, Gmail’s sentence completion feature, or personalized ads, but the application of ML in materials discovery is evidence that data can be better leveraged to create physical solutions as well.