Nobel for AI: A matter of time?

Can an Artificial Intelligence (AI) be awarded the Nobel Prize for key scientific research? This is not a frivolous or philosophical question. In 2022 alone, we have seen at least two AI-driven pieces of research, which could qualify for the most prestigious award in science.

In an earlier era, one would, of course, have cited the flesh-and-blood scientists who did computer-assisted work. But in both the cases in question, the AI was essentially self-taught — actually, it was arguably the same AI that did both pieces of research in widely differing fields.

In February 2022, DeepMind, the British AI research outfit, which is an Alphabet subsidiary, was asked to help look at a critical problem in nuclear fusion. Fusion occurs naturally in the sun and other stars. Extremely hot hydrogen is squeezed at extreme pressures to turn into helium, with the excess particles being converted to energy. We find it hard to create such conditions of extreme pressure and temperatures on Earth.
 

Tokamaks are doughnut-shaped chambers, designed by Andrei Sakharov back in the 1950s, which are used in nuclear fusion. Plasma (hot ionised gases) is contained within a tokamak and crunched by magnetic manipulation to initiate nuclear fusion.

“Hot” in this case means 120 million degrees Celsius — much hotter than the sun. There is no known material, which can contain anything at such temperatures without itself melting, and if that plasma touches the chamber walls, it could be catastrophic. But plasma is by definition, electrically positive (since electrons have been stripped off). So it can be contained by strong electro-magnetic fields.

Only since it’s gaseous, those magnetic fields need to be reconfigured many times per second, to keep the plasma contained. Doing this efficiently, while using minimum energy to run the magnets, and shaping the field to squeeze plasma in ways that yield maximum fusion, is very hard. The research by DeepMind’s AI was laid out in the paper “Magnetic control of tokamak plasmas through deep reinforcement learning” (this has human authors). The AI had worked out far more efficient ways to shape the fields and hold plasma than any prior research.

The other piece of top notch AI-driven research was released last week. This involves protein-folding, which is an insanely complicated mathematical problem. There are roughly 200 million known proteins, each composed of about 20 amino acids. Proteins consist of chains of 300-odd amino acids or more strung together in various ways. Any protein can be folded many different ways. The biochemical actions of proteins depend on the way they are folded — this drives interactions since chemistry depends on physical proximity. If a protein is misfolded, horrible things occur such as congenital diseases.

Understanding how proteins are folded, or likely to be folded is, therefore, critically important. Protein folding is guesswork, or rather, it used to be guesswork driven by combinatorial maths and experimental observation. Human researchers using experimental methods and cryo-electron microscopy have often taken years to figure out the structure of a single protein.

Circa 2018, DeepMind’s AlphaFold algorithm started working on protein-folding. This “Alpha” algorithm is the same self-learning system that taught itself to play superb chess and Go.

In 2018, AlphaFold beat every other protein-guessing program in the 13th Critical Assessment of Protein Structure Prediction, an annual contest where programs try to guess structures of proteins. It won the 14th edition as well.

In 2021, AlphaFold was open-sourced. The source code was released along with a paper that outlined the detailed methodology of guessing protein folding. On July 28, DeepMind released a compendium of the structures of 200 million proteins. The AI seems to have more or less solved the protein-folding problem.

This is a huge step forward and it could bring untold future benefits. Any human team, which figured this out, would have been a more or less automatic recipient of the Nobel. So, to return to the original question — should self-learning AI be eligible for science awards?