Imagine You Could Make Anything

What would you make? What would you make it out of? This weeks deep dive is going to explore a breathrough, announced this week, by Google DeepMind, about a major breakthrough in materials science. To be honest, not something that has ever been at the forefront of my mind before. But it is now…

Unlocking Millions of New Materials with Artificial Intelligence

For all of human history, we’ve relied on materials found in nature – stone, wood, metal – to build our world. But in recent decades, scientists have begun deliberately creating new materials in labs, designing them with specific properties needed for advanced technologies like batteries, solar panels, and computing chips.

The challenge is that searching for the best new materials by randomly trying combinations is incredibly slow. There are just too many potential materials to explore. So researchers at Google DeepMind tried a radically new approach – using artificial intelligence to guide the search.

They built a deep neural network, a type of AI model that learns from data, and trained it on a database of known stable materials. The model acts like an apprentice looking over an expert’s shoulder, learning patterns about what combinations of elements form useful materials.

Armed with this knowledge, the AI could start predicting the stability of hypothetical new material candidates. The researchers generated over a billion new material ideas, drawing on enormous computing power. The AI filtered down the candidates, predicting which had the highest chance of being stable. Only the most promising were synthesised and tested in the lab using a technique called density functional theory.

When a prediction was accurate, the new material was added back into the dataset, allowing the AI to learn from its successes and failures. With each iteration the model improved, on a quest to master the alchemy of materials science.

After several rounds of prediction, synthesis, and learning, the AI had discovered over 2 million completely new stable materials – a 10x expansion in known materials. The researchers were stunned that the AI could generalise so well, even accurately predicting materials made of 5 or 6 different elements. Such complexity normally requires expert human intuition.

Among the open questions is how many of these materials can be experimentally constructed. But the sheer scale of compounds found demonstrates AI’s potential to accelerate discoveries. And with each new material added to its knowledge, the model becomes smarter, like a scientist who has read millions more research papers than any human could digest in a lifetime.

One researcher explained that they built interatomic potentials, mini AI models that simulate the physics of how collections of atoms interact. After pre-training on the 2 million materials discovered, these potentials could accurately predict properties of new materials they’d never seen before. They were even useful in applications like identifying promising solid electrolyte materials for better batteries.

So beyond just finding stable materials, the AI learned reusable knowledge to bootstrap new investigations. With these virtual laboratories in hand, the scientists anticipate designing specialised materials for carbon capture, clean energy, and more. Imagine being able to design a material that eats and stores carbon like a tree.

The impacts of the research extend into the nature of science itself. Scientists often seek fundamental explanatory principles that fit past observations while predicting future discoveries. But machine learning turns this approach upside down, learning patterns from data first and deriving human-interpretable theories later.

The researchers believe AI can transcend the limitations of human intuition. And we’ve merely scratched the surface of what powerful models coupled with experimentation can unlock. Computational discovery schemes are already transforming chemistry, physics, biology and medicine.

One day, we may view the millions of known natural materials as quaint compared to the infinite library of bespoke designs conceived by intelligence – both biological and artificial. Through this fusion of mind and machine, everything from smartphones to space colonies may one day be crafted atom by atom.

To read more on this, here is a link to my post on X as well as the original source paper published in Nature here.